Dream & Passion

WSJ Setup

2012-02-01T18:34:00.001+08:00

WSJ SI84, 7138 sentences, (removing 410's duplication in the original data)

WSJ SI284, 37416 sentences, (removing 410's duplication in the data set)

speaker 410 has contributed two sessions of recordings, thus has twice the number of sentences as other speakers. During the processing, we should remove 11-2.1/wsj0/si_tr_s/401 to avoid too much data for a single speaker.

The 5K and 20K test sets for WSJ0 are Nov.92 test sets.

The 64K test sets of WSJ1 is the commonly mentioned Nov.93 test set.

http://books.google.com.sg/books?id=WJeAWYFa0i0C&pg=PA127&lpg=PA127&dq=WSJ+SI284+number+sentences&source=bl&ots=I2laYM2emM&sig=plZXYseK9n6umvawaWZ7vhOrhcw&hl=en&sa=X&ei=ueUoT5P5Jo-HrAfQ4oG9AQ&ved=0CGYQ6AEwBw

Posted via email from Troy's posterous

[HTK] Chinese Encoding

2011-11-30T18:54:00.001+08:00

修改HTK源码 HParse，HVite部分，使其支持中文

2010-03-24 12:05

From: http://hi.baidu.com/cbyhit2008/blog/item/e642d6c7c16bd4179c163d5b.html

利用HTK工具包进行语音识别建模时，遇到任务语法中存在中文时候，无法生成对应的底层网络，这样就需要对HTK源码的部分内容进行修改，以下是我对HTK源码HParse及HVite部分内容改动记录，希望对有需要的人有帮助！自己也做个备份！
添加下面函数
static int IsSpace(char c)
{
if ((c == 0x09) ||( c == 0x0D) || (c == ' ' ))
return 1;
return 0;
}
修改下面的函数
static void PGetSym(void)
{
..../////////////
+++while ( !IsSpace(ch) || (ch=='/' && inlyne[curpos]=='*') ) //isspace((int) ch)
{
+++    if (!IsSpace(ch) || isspace((int) ch)) /* skip space */
PGetCh();
else {            /* skip comment */
PGetCh(); PGetCh();
while (!(ch=='*' && inlyne[curpos]=='/')) PGetCh();
PGetCh(); PGetCh();
}
}
..../////////////以下部分代码为做修改
}

static void PGetIdent(void)
{
int i=0;
Ident id;

do {
if (ch==ESCAPE) PGetCh();
if (i<MAXIDENT) id[i++]=ch;
PGetCh();
+++ } while (!IsSpace(ch)&& ch!='{' && ch!='}' && ch!='[' && ch!=']' &&//!isspace( (int)ch)
ch!='<' && ch!='>' && ch!='(' && ch!=')' && ch!='=' &&
ch!=';' && ch!='|' && ch!='/' && ch!='%');
id[i]='\0';
ident = GetLabId(id,TRUE);
}

ReturnStatus WriteOneLattice(Lattice *lat,FILE *file,LatFormat format)
{
...///////////////////////////////
else if (ln->word!=NULL) {
fprintf(file,"W=%-19s ",ln->word->wordName->name);//
//   ReWriteString(ln->word->wordName->name,注释掉
//                NULL,ESCAPE_CHAR));
...////////////////////////////////
}
这样在生产的底层网络中就可以看到汉字，而不是汉字编码了。下面是我测试的一个简单例子：
这是taskgram中的内容
$word = 好
| 浩
| 尼
| 你;
( START_SIL ([sil] )(<$word>)( [sil]) END_SIL )
没有修改HParse生产的网络
VERSION=1.0
N=11   L=22
I=0    W=END_SIL
I=1    W=sil
I=2    W=\304\343
I=3    W=!NULL
I=4    W=\304\341
I=5    W=\272\306
I=6    W=\272\303
I=7    W=sil
I=8    W=START_SIL
I=9    W=!NULL
I=10   W=!NULL
J=0     S=1    E=0
J=1     S=3    E=0
J=2     S=3    E=1
J=3     S=3    E=2
J=4     S=7    E=2
J=5     S=8    E=2
J=6     S=2    E=3
J=7     S=4    E=3
J=8     S=5    E=3
J=9     S=6    E=3
J=10    S=3    E=4
J=11    S=7    E=4
J=12    S=8    E=4
J=13    S=3    E=5
J=14    S=7    E=5
J=15    S=8    E=5
J=16    S=3    E=6
J=17    S=7    E=6
J=18    S=8    E=6
J=19    S=8    E=7
J=20    S=10   E=8
J=21    S=0    E=9
修改后的网络
VERSION=1.0
N=11   L=22
I=0    W=END_SIL
I=1    W=sil
I=2    W=你
I=3    W=!NULL
I=4    W=尼
I=5    W=浩
I=6    W=好
I=7    W=sil
I=8    W=START_SIL
I=9    W=!NULL
I=10   W=!NULL
J=0     S=1    E=0
J=1     S=3    E=0
J=2     S=3    E=1
J=3     S=3    E=2
J=4     S=7    E=2
J=5     S=8    E=2
J=6     S=2    E=3
J=7     S=4    E=3
J=8     S=5    E=3
J=9     S=6    E=3
J=10    S=3    E=4
J=11    S=7    E=4
J=12    S=8    E=4
J=13    S=3    E=5
J=14    S=7    E=5
J=15    S=8    E=5
J=16    S=3    E=6
J=17    S=7    E=6
J=18    S=8    E=6
J=19    S=8    E=7
J=20    S=10   E=8
J=21    S=0    E=9
至于HVite部分，我找了近一下午，总算找到改的地方了，修改HSheel.c 中WriteString函数
n=*p;
fputc(n,f);
//   fputc(ESCAPE_CHAR,f);
// fputc(((n/64)%8)+'0',f);fputc(((n/8)%8)+'0',f);fputc((n%8)+'0',f);
我将相应的位置给注释上了，并将字符之间输出到文件中，这样在结果文件中就可以看到中文了～～

Posted via email from Troy's posterous

[HTK] Chinese encoding

2011-11-30T18:52:00.001+08:00

HTK could directly read in the "gbk" encoded MLF or dictionary etc. files. Actually, it could read any kine of encoded file. In HTK, what it does is to read in every byte (char type) and when print them out, each byte is write out in the form of "\abc", which abc is the octal representation of the byte number(=a*64+b*8+c).

Thus to convert the HTK generated files back to the readable characters, we need following steps:

1) convert the HTK octal representation of byte values to byte array

2) decode the byte array with corresponding encoding, (e.g. for Chinese, we could use "gbk")

Following is the code I used to convert the HTK generated MLF to readable "gbk" encoded MLF file:

import string, codecs

fin=open('vom_utt_wlab.mlf')

fout=codecs.open('vom_utt_wlab.gbk.mlf', encoding='gbk', mode='w')

while True:

sr=fin.readline()

if sr=='':break

sr=sr.strip()

if sr.endswith('.lab"'):

print >>fout, sr

while True:

sr=(fin.readline()).strip()

if sr=='.':break

if sr.startswith('\\'):

lst=(sr.strip('\\')).split('\\') # get the list of octal representation of each byte

bins=bytearray()

for itm in lst:

val=0

for ii in range(3): # each octal number will have exactly 3 numbers, i.e. of the form \nnn

val=val*8

val=val+int(itm[ii])

bins.append(val)

print >>fout, bins.decode('gbk')

else:

print >>fout, sr

print >>fout, '.'

else:

print >>fout, sr

fin.close()

fout.close()

Posted via email from Troy's posterous

[HTK] Increase HTK feature dimension limit

2011-11-24T17:21:00.001+08:00

In the HTK feature file, there is a header file specify the basic information of the parameters.

HTK format files consist of a contiguous sequence of samples preceded by a header. Each sample is a vector of either 2-byte integers or 4-byte floats. 2-byte integers are used for compressed forms as described below and for vector quantised data as described later in section 5.11. HTK format data files can also be used to store speech waveforms as described in section 5.8.

The HTK file format header is 12 bytes long and contains the following data

nSamples                - number of samples in file (4-byte integer)

sampPeriod - sample period in 100ns units (4-byte integer)

sampSize - number of bytes per sample (2-byte integer)

parmKind - a code indicating the sample kind (2-byte integer)

http://blush.ee.columbia.edu/doc/HTKBook21/node58.html

From the above specification, the sampSize is short integer, thus the maximum value for sampSize is 32768. For uncompressed data, the maximum dimension for each sample is thus 32768/4=8192. However, usually even just 1000+ D feature will cause the HTK tools to generate following errors:

OpenParmChannel: cannot read HTK Header in File

The reason is that in the function ReadHTKHeader of the file HWave.c, there is check for the sampSize value:

if (hdr.sampSize <= 0 || hdr.sampSize > 5000 || hdr.nSamples <= 0 ||

hdr.sampPeriod <= 0 || hdr.sampPeriod > 1000000)

return FALSE;

That's to say, in HTK the dimension of the feature vector is limited by this check instead of data type specified in the header format. In the standard version of HTK, at most 1250D feature could be used. To increase the limit, what we need to do is to change the number 5000, but do remember sampSize is short integer, changing to any value larger than 32768 would be useless.

The code at about line 1427 of the file HTKLib/HWave.c.

Posted via email from Troy's posterous

Solving iMessage or FaceTime waiting for activation problem

2011-11-08T19:44:00.001+08:00

My problem is even further than simply "waiting for activation". After change the SIM card of my iphone, the iMessages sent out are all under the old number ....

However, the number under the iMessage setting is unchangeable. To change it, you need to turn off both the iMessage and FaceTime and then reactivate them. The problem comes when reactivating them, the "waiting for activation" messages last for hours without giving any hints what are going on.

Here comes the "waiting" problem as many people encountered. To solve it, Simply speaking, you need to save your own contact information in the Contacts and in the general settings set my info to the correct contact.

From: https://discussions.apple.com/thread/3390466?start=0&tstart=0

Hi all....I have an iPhone4 with IOS5 and was getting the waiting for activation message for imessages. I was also finding that people I was iMessaging were getting messages from my email address (aka apple ID) rather than my mobile number which was confusing them. I found following adnanfarooqui's entry above and the iphoneism site very useful and it fixed my iPhone even though it was more directed at Facetime.

All I did was simply go into my own details in CONTACTS and ensure my own mobile # was set up as 'iphone' in contacts. Then I went into SETTINGS>MAIL,CONTACTS,CALENDAR>CONTACTS>MY INFO and selected myself. What this does, I suspect, is tells the iphone who I am as the owner along with the mobile #. Then I went back into SETTINGS>iMESSAGES and found that the 'waiting for activation' message had changed to the normal message with the 'learn more' link included. Then I went down to RECEIVE AT, got into my Apple ID and signed out of that (which I suspect was what my imessages recipients were seeing), which then allowed me to select my own mobile # as the receive at number. This therefore also fixed the issues I was having with recipients of my imessages getting my email address as the 'from' details as opposed to my mobile # and not matching their own CONTACTS. Hope that was helpful......

Posted via email from Troy's posterous

Developing a Static Library and Incorporating It in Your Application in Xcode4

2011-11-08T16:37:00.001+08:00

Developing a Static Library and Incorporating It in Your Application

From: http://developer.apple.com/library/ios/#documentation/Xcode/Conceptual/ios_development_workflow/910-A-Developing_a_Static_Library_and_Incorporating_It_in_Your_Application/archiving_an_application_that_uses_a_static_library.html

When you need to develop a static library to use in an application and you have to have separate projects for each product, you can use a workspace to contain both the static library project and the application project. If you do, ensure that you configure the projects in the workspace as described here:

In the target that builds the static library, ensure that:
- The exported headers are in the Project group in the Copy Headers build phase.
- The Skip Install build setting is set to Yes.
In the target that builds the application, ensure that:
- The User Header Search Paths build setting is set to the recursive absolute path of a directory under which the static library’s header files are stored.
  
  Important: If you move your static library project directory to a different location in your file system, you must update the value of the User Header Search Paths build setting to reflect the new location of the static library’s header files.
- The Always Search User Paths build setting is set to Yes.
- The Skip Install build setting is set to No.
In the scheme that builds the application, ensure the scheme also builds the static library for archiving.

Content specifications: This content is written for Xcode 4.0.2 and iOS SDK 4.3.

Posted via email from Troy's posterous

Tutorial: Code Sharing Via Static Libraries And Cross-Project References

2011-11-07T22:01:00.001+08:00

Tutorial: Code Sharing Via Static Libraries And Cross-Project References

CLINT HARRIS on 01. APR, 2009 in RESOURCES, TIPS, TUTORIALS

From: http://mobileorchard.com/code-sharing-via-static-libraries-and-cross-project-references/

Guest author Clint Harris (Profile) is an independent software consultant with experience ranging from enterprise web app work to custom iPhone app development. He currently lives in Brooklyn, New York.

Finding an elegant way to reuse and share code (i.e., libraries) across separate iPhone applications can be a bit tricky, especially considering Apple’s restrictions on dynamic library linking and customFrameworks.

Most people agree that the best way to re-use code is to use static libraries. This tutorial builds on that solution, showing how your Xcode project can reference a second Xcode project — one which is used to build a static library.

This approach allows you to automatically build that static library with the rest of your app, using your current build configuration (e.g., debug, release, etc.) and avoid pre-building several versions of the library separately (where each version was built for a specific environment/configuration).

Wanted: An Elegant Way To Share Code Across Projects

If you want to reuse/share code across different iPhone applications, you only have two options that I’m aware of:

Copy all of the source code from the “shared” library into your own project
Keep the shared library code in a separate Xcode project and use it to build static libraries (e.g., libSomeLibrary.a, also referred to as “archive files”) that can be referenced by your project and used via static linking.

The first option, copying the files, should be avoided when possible since it’s inherently redundant and contrary to the goal of keeping “common code” modular and atomic.

It’s a much better idea to put the code in a static library (see since, as mentioned in the introduction, dynamic linking to custom libraries/frameworks isn’t allowed by Apple when it comes to iPhone apps.

For instructions on creating a static library from your code see this tutorial on the Stormy Productions blog.

We’ve established that the second option is preferable, but there’s a catch: you’ll need to build and distribute multiple versions of the static library–one for each runtime environment and build configuration. For example, you would need to build both “release” and “debug” versions of the library for the Simulator, as well as other pairs for the iPhone or iPod device itself.

So, how can we avoid manually pre-building and managing separate .a files?

Solution: Static Libraries Built On-Demand Via Xcode Cross-Project References

The trick to avoid pre-building static libraries for each environment is to use an Xcode “cross-project reference” so that those libraries are built dynamically (i.e., when you build your own app) using your app’s current build configuration.

This approach allows you to both reuse shared source code and avoid the headache of managing multiple versions of the library. Here’s how it works at a high level:

The shared code lives in its own Xcode project that, when built, results in one or more static libraries.
You create an Xcode environment variable with a path to the directory that contains the static library’s *.xcodeproj file.
All iPhone apps that need the static library will use the aforementioned environment variable toreference the library’s Xcode project, including any static library in that project and the related header files.
Each time you build your project for a specific configuration/runtime environment, the shared project library will also be built for that config/environment–if it hasn’t already–and linked with your executable.

In addition to solving the main problem (reusing code and avoiding management of multiple library versions), there are a couple of nice benefits to this strategy:

First, if you make changes to the shared library source code, those changes will immediately be included the next time you build your own project (via the cross-project reference).

Second, you can modify the Xcode environment variable to point to different versions of any project. For example, you might have separate directories for “somelibrary-1.0″ and “somelibrary-2.0″; as you’ll see in the detailed solution instructions, it’s easy to modify the environment variable and switch your project to a different version of “somelibrary.”

Implementing Cross-Project References

The instructions for setting up cross-project references to shared static libraries can be split into two parts:

Part 1: Global Xcode Settings
Part 2: Project-Specific Settings

I’ll be using an example in the instructions to help illustrate things. A suitable example would be an application that needs to use a shared static library from a separate project. In this case, I’ll use a sample iPhone app called “

Posted via email from Troy's posterous

HDecode in lattice rescoring mode

2011-11-04T12:57:00.001+08:00

When training NN, the training process is usually controlled by the frame accuracy (or frame error rate). However, it is not directly related to the speech recognition performance, i.e. PER or WER.

One way is to do decoding after each time the network weights are updated. For phoneme recognition, it is fine, as the decoding doesn't take too much time. When coming to word recognition, the decoding is quite time consuming. To speed up, one possible way is to using lattices instead a full decoding.

Invoke HDecode with "-w" without language model, it will run in lattice rescoring mode (of course, you need set the input lattices parameters). But where to get the lattices? My setup is as follows:

1) Using HDecode and HMM system to generate lattices using bigram LM ( can also use higher order LM);

2) Using HLRescore to prune the lattices to word networks ( with -m f/b, and save the new lattices with -w);

3) Using HDecode to rescore the lattices with new acoustic model or new posterior features ( in NN case).

Posted via email from Troy's posterous

[HTK] CreateHMM: multiple use of logical HMM name sp

2011-10-12T14:15:00.001+08:00

When expanding the monphone system to triphone system, I encountered this problem.

From the page : http://hts.sp.nitech.ac.jp/hts-users/spool/2009/msg00271.html, it suggests to check the mktri.hed file, where I found the expanding of sp as triphones.

Thus the problem is in the process of generating the mktri.hed file, which is done using:

perl maketrihed [mono_list] triphn_seen.lst

The correct way of doing so is:

perl maketrihed [mono_no_sp_no_sil_list] triphn_seen.lst

Posted via email from Troy's posterous

PCA whitening

2011-09-29T11:53:00.001+08:00

I have encountered this phrase "PCA whitening" many times, all I thought about it was just a simple PCA. However, it is not!

The Whitening is actually PCA + scaling!!!

Also there is a good implementation explanation here : http://ufldl.stanford.edu/wiki/index.php/Implementing_PCA/Whitening

Implementing PCA_Whitening - Ufldl.pdf Download this file

Posted via email from Troy's posterous

PCA again

2011-09-29T11:36:00.001+08:00

PCA is a simple but rather popular techniques for feature processing and is used widely as some initial processing steps involved in many applications. However, every time when I need to do PCA, I have to find some references :<.

Thus again, to put some information for my own reference. The figure is from the attached slides I encountered today, quite clear.

3 - Feature extraction.pdf Download this file

Posted via email from Troy's posterous

Unity3D Resources

2011-09-03T10:40:00.001+08:00

The original post is from: http://www.virtualgamelab.com/2/post/2010/04/unity-resources.html
This is just for my own reference!

As promised, this blog entry is going to give an overview of existing Unity resources. I have tried to cover a wide variety of video and text-based tutorials, extensions for Unity and links to most example projects and documentation on the web. I have tried to categorize the resources and also present them in an order that should get you started in no time. Furthermore, I have spent a substantial amount of hours either absorbing these tutorials/docs completely or at least scanning through them (I haven't bought any paid tutorials except Will Goldstone's book). Thus, I believe all of these links should lead you to some form of valuable information about Unity. Thanks to all the authors for doing such an incredible job. Without further ado, my take on summarizing resources for Unity.

Unity website: http://unity3d.com/

Which license do you need? http://unity3d.com/unity/licenses

Download the free version of Unity to get started: http://unity3d.com/unity/download/

If you have no previous experience with Unity, start with these six video tutorials which give a quick overview of the Unity interface and some important features
http://unity3d.com/support/documentation/video/

Continue with a more in-depth text-based walk through of very basic Unity functionality and work flow http://unity3d.com/support/documentation/Manual/Unity%20Basics.html

To get you started with scripting, have a look at the following PDF document. It was written for an older version of Unity, but still covers relevant aspects of scripting with JavaScript. (2 hours to complete, no previous JavaScript knowledge required)
http://docwiki.unity3d.com/uploads/Main/Scripting%20Tutorial.pdf

Unity features three scripting languages; JavaScript, C#, and Boo. Depending on your scripting language of choice, you might want to choose different tutorials to get started with Unity. Each tutorial link will also mention which scripting language is used. Most Unity tutorials available on the web are using JavaScript. Notice that JavaScript is commonly used for web applications. Unforunately, books which cover JavaScript are normally of little use for Unity's implementation of JavaScript. But worry not, a vast amount of tutorials will follow...

If you want to familiarize yourself with Unity's functionality more, browse through Unity's manual. You can skip the “Basics” section as we already went through this (see above).
http://unity3d.com/support/documentation/Manual/index.html

For a printable version of the 600+ pages manual, seehttp://unity3d.com/support/documentation/printable.html

If you rather prefer to jump into the action, skip the manual altogether and have a look at the three books which have been published on Game Development with Unity (one recently added on 06/10/11).
1) Author Will Goldstone guides you through a complete project and introduces most of Unity's functionality (using JavaScript). The book is available as print and digital version and all needed assets and resources can be downloaded with the book. It's been a helpful investment from my point of view.
http://www.packtpub.com/unity-game-development-essentials/book
This forum thread discusses the book and its content.
http://forum.unity3d.com/viewtopic.php?t=29041

2) Ryan Henson Creighton more recently published "Unity 3D Game Development by Example Beginner's Guide". Find out more about the book here: https://www.packtpub.com/unity-3d-game-development-by-example-beginners-guide/book
and the forum thread - http://forum.unity3d.com/threads/61349-Unity-3D-Game-Development-by-Example

3) Craig Stevenson and Simon Quiq (publisher Deep Pixel) released "Unity 3 Blueprints: A practical guide to Indie games development". Their website provides all the art assests and code to create four classic games from scratch (Match the Pairs, Top-Down Shooter, Tower Defense, Marble Madness).
Amazon link

To get up-to-date on Unity 3 functionality, check out the following official documentation:

Unity 3 - What's new? http://unity3d.com/unity/whats-new/unity-3

- Beast lightmapping Basics - http://unity3d.com/support/documentation/Manual/Lightmapping.html
- Beast Lightmapping In-Depth - http://unity3d.com/support/documentation/Components/class-LightMapping.html
- Tree Creator - http://unity3d.com/support/documentation/Components/class-Tree.html
- Umbra Occlusion Culling - http://unity3d.com/support/documentation/Manual/Occlusion%20Culling.html

You are now faced with the choice of tackling example projects provided by Unity Technologies or jump straight into user-generated tutorials. As the example projects are quite complex, I would suggest working through some video tutorials first. Nonetheless, here are the links for Unity's official example projects. They can be taken completely apart, reverse-engineered and reused for own projects.

3D Platform Game
http://unity3d.com/support/resources/tutorials/3d-platform-game
2D Platform Game
http://unity3d.com/support/resources/tutorials/2d-gameplay-tutorial
Iphone tutorials and more example projects
http://unity3d.com/support/resources/tutorials/

More Unity Example projects
http://unity3d.com/support/resources/example-projects/

Brand New Car Tutorial by Unity Technologies
http://unity3d.com/support/resources/tutorials/car-tutorial

Now, let's head on to TUTORIALS. This list starts with mostly basic and general tutorials at the top and lists more specific tutorials at the end.

Will Goldstone , author of the Unity Game Development book, worked on a series of helpful video tutorials here (using JavaScript): http://learnunity3d.com/ (also http://learnmesilly.com/).
More recently by Will Goldstone: http://www.unity3dstudent.com/

TornadoTwins Video Tutorials
They show step by step how to create a simple game using Javascript.
http://www.youtube.com/user/TornadoTwins

Walker Boys Studio - Unity Tutorials from the Guildhall at SMU (an extensive list of tutorials with more on the way)
http://walkerboystudio.com/html/unity_training___free__.html

CannedMushroom Video Tutorials (Unity and other software)
This is a series of projects intended for 2-hour self-instructed lessons using JavaScript
http://www.youtube.com/user/cannedmushrooms

Unity Jumpstart
Proof of Concept Game to learn creating your own game from the ground up (JavaScript)
http://forum.unity3d.com/viewtopic.php?t=28433
http://www.unityjumpstart.com/ProofOfConcept_1/

3dbuzz.com
Series of Unity tutorials in C# (among many other tutorials)
registration required (free)
website - http://www.3dbuzz.com/vbforum/sv_home.php
forum thread - http://forum.unity3d.com/threads/37093-Over-7-hours-of-Unity-Training-Videos-Download-them-now!

BurgZergArcade - Unity Tutorials
Hack&Slash Tutorial using C# and plenty of other information and useful tutorials
website - http://www.burgzergarcade.com/

UnityScript Basics (Scripting Basics for Noobs)
If you're new to scripting, read up on this good introduction to scripting in Unity. Lots of analogies are provided which makes it really easy to understand. The details are explained for C#, but the introduction is great for any scripting language.
website - http://www.unityscript.com/lessons1/basics.php

UnityLabs
Text-based tutorials with screenshots covering a wide range of topics incl. scripting (using JavaScript), basic introduction to unity, character controllers, and user interfaces
http://www.unitylabs.net/tutorials/unity-beginners

InfiniteAmmo Tutorial
General Introduction to Unity (3 parts so far)
Some scripting covered using Javascript (e.g. movement controls)
http://infiniteammo.ca/blog/impromptu-unity3d-intro-tutorial/

Technicat
Introduction to Game Development with Unity including Workflow, Scripting, GUI, Version Control, etc.
http://technicat.com/games/unity.html

Advanced Media Lab at North Carolina State University
http://vimeo.com/channels/ncsuunity

GearTech Games
4 Videos on how to work through a project (and more videos)

workflow between Modo and Unity
lightmapping

http://www.geartechgames.com/GearTech_Games/Beginning_a_Project/Beginning_a_Project.html

Making Sense of Unity
These video tutorials provide more indepth coverage of unity scripting using Javascript (not meant to be introductory material)

lecture style explanations of concepts
more entertaining and not like most other screen-capture tutorials :)

http://www.youtube.com/etphonemyhome

IO Development Diary
This series of video tutorials (some paid, some free) follows the development of a Space Shooter and covers a Modo-Unity workflow and C# scripting.
forum thread - http://forum.unity3d.com/viewtopic.php?t=55762
website - http://fallingbrickwork.fearedfuturegames.com/

Virtual Autonomy
Guide for working with Google SketchUp and Unity
(also shaders and multi-user environments)
text-based with screenshots
http://virtualautonomy.com/blog/index.php/tutorials/

Robotduck - Blog
The blog of this Unity user provides tips on Unity scripting and functionality and showcases some of his projects
http://robotduck.wordpress.com/

Ethical Games
Unity Tutorials for Flash Developers
http://ethicalgames.wordpress.com/2009/01/14/unity-for-flash-developers-tutorial-1/

Car Tutorial
Physics Setup for a car, AI for driving around racetrack
http://www.gotow.net/andrew/blog/?page_id=78

Terrain Tutorial
seven videos on how to create terrains in Unity from heightmaps to finish
http://vimeo.com/album/156452

In-Depth Terrain Tutorial
text-based with screenshots
http://www.unifycommunity.com/wiki/index.php?title=Terrain_tutorial
forum thread
http://forum.unity3d.com/viewtopic.php?t=16941

Paul Bourke -Unity for stereoscopic display
text-based with screenshots
http://local.wasp.uwa.edu.au/~pbourke/miscellaneous/stereographics/Unitystereo/

Paul Bourke – Unity and Idome
text-based with screenshots
http://local.wasp.uwa.edu.au/~pbourke/miscellaneous/domemirror/UnityiDome/

RENCI – Unity for Dome projections
http://unc.renci.org/rencinews/the-dome-corner-006

Official Unity Tech. Tutorial for the Animation View (NEW) 3-part-series of video tutorials
http://unity3d.com/support/resources/tutorials/video-animation-view

PAID TUTORIALS

VTC - Unity tutorial using JavaScript (subscription based content on vtc.com)
http://www.vtc.com/products/Introduction-to-Game-Development-Using-Unity-3D-Tutorials.htm
Unity-Tutorials (mostly paid and some free tutorials)
http://www.unity-tutorials.com/

Design3.com (subscription based tutorials, Unity-related content among other software like Maya, 3DS Max, Photoshop)
http://www.design3.com/by-software/unity

GENERAL RESOURCES

Once you went through some or all of these tutorials, you should be creating your own content in no time. Next, I'll provide an overview of general resources for Unity. Whenever you need to find some information about Unity, scripting, projects, collaborations or anything Unity-related, check these links out:

Searching for resources - The All-In-One Unity Reference Search
http://www.google.com/cse/home?cx=001712401338047450041:csfhqk-trfa
credit goes to Robotduck for providing this link to the public; tremendous time saver (seehttp://robotduck.wordpress.com/)

Unity Scripting Reference
http://unity3d.com/support/documentation/ScriptReference/index.html

Unity Component Reference
each available Component described in detail
http://unity3d.com/support/documentation/Components/index.html

Unity Wiki (UnifyCommunity)
http://www.unifycommunity.com/wiki/index.php?title=Main_Page

Unity Answers
Invaluable when you have specific questions about Unity or Scripting
http://answers.unity3d.com/
This link should also get you started on learning Unity
http://answers.unity3d.com/search?q=learning+tutorial

Unity Forum
http://forum.unity3d.com/
http://forum.unity3d.com/viewtopic.php?t=30633

Unity Feedback
feature requests go here
http://feedback.unity3d.com

Unity IRC
Point your favorite IRC client to irc.freenode.net and join #unity3d to chat in real time.

Overview of Unity Resources
http://www.reddit.com/r/unity_tutorials/

Overview of Unity blogs
http://www.reddit.com/r/unity_blogs/

TOOLS AND EXTENSIONS

Lastly, I want to list extensions and tools which can make your life as a Unity developer easier.

Unity Extensions

Terrain Toolkit
External Lightmapping Tool
Locomotion System
Explosion Framework
Head Look Controller

http://unity3d.com/support/resources/unity-extensions/
overview of some extensions – Unity youtube and vimeo channels
http://www.youtube.com/user/Unity3D
http://www.vimeo.com/unity3d/videos

Visual C# Express- free IDE for your C# development
http://www.microsoft.com/express/Downloads/#2008-Visual-CS

3DAttack - Tools and Home of Unity Creative Magazine
First Person Shooter Developer Kit
forum thread - http://forum.unity3d.com/viewtopic.php?t=53136
website - http://www.3dattack.us/3DAttack/News.html

AngryAnt
One of Unity's developers provides projects to extend Unity's functionality (e.g. pathfinding and AI)
http://angryant.com/
forum thread Path - http://forum.unity3d.com/threads/19341-Pathfinding-in-unity-First-Path-1.0-tutorial-now-available!

sturestone's A* Pathfinding (currently version 2.9)
forum thread - http://forum.unity3d.com/threads/67417-A*-Pathfinding-2.9-Is-Released-(Unity-3-Compatible)

Weltenbauer
Augmented Reality / Webcam Input
http://produktion.weltenbauer.com/#/5
forum threads:
http://forum.unity3d.com/viewtopic.php?t=32613&highlight=toolkit (Webcam Toolkit)
http://forum.unity3d.com/viewtopic.php?t=11889&highlight=toolkit (ARToolkit Extension)
http://forum.unity3d.com/viewtopic.php?t=32669&highlight=toolkit (UnityAR)

SeeingMachines FaceAPI / VisionBlaster – Head Tracking in Unity (purchase required)
http://visionblaster.com/cms/index.php?option=com_content&view=article&id=48&Itemid=62&lang=en
and http://www.seeingmachines.com/product/faceapi/downloads/

Mostly Tigerproof – Using Google Analytics and Unity to track game stats
This is a blog entry about Google Analytics and Unity
http://blog.mostlytigerproof.com/2009/10/06/gathering-statistics-using-google-analytics-and-unity-3d/

Antares Project - Extensive Set of Tools to extend the Unity Editor - Open Source
Also available: Antares.dll (free for non-commercial work)
forum thread - http://forum.unity3d.com/viewtopic.php?t=50843

Antares Deformator - Deform your meshes (Beta Version)
forum thread - http://forum.unity3d.com/viewtopic.php?t=54290

UniWii – WiiMote implementation
http://www.unifycommunity.com/wiki/index.php?title=UniWii
forum thread
http://forum.unity3d.com/viewtopic.php?t=7284

Unity Terrain Tools - EasyRoads (purchase required)
http://unityterraintools.com/downloads.php
forum thread
http://forum.unity3d.com/viewtopic.php?t=23519

Six Times Nothing - Road/Path Tool and River Tool http://www.sixtimesnothing.com/

Dastardly Banana - FPS Weapon Tool, Radar examplehttp://dastardlybanana.com/Downloads.htm

Starscene Software - Tools, Games and Utilities for Unity (purchase required for utilities)
http://www.starscenesoftware.com
e.g. Vectrosity - Line Drawing Tool
http://starscenesoftware.com/vectrosity.html
Fractscape - Terrain Tool
Stitchscape - Stitch multiple terrains together

GUIX - visual Menu/GUI builder (purchase required)http://www.ennanzus.com/developer/GUIX/buy.htm

EZ Game Saver - saving tool (purchase required)
http://www.anbsoft.com/middleware/ezs/index.htm
Note that I will cover saving to text file in a later blog

Fire Tool - Realistically spreading fire (on hold for the moment)
http://forum.unity3d.com/viewtopic.php?t=49588

Decal Framework - Easily place decals in your scene
forum thread - http://forum.unity3d.com/viewtopic.php?t=54613

Visual Logic Editor by NeoDrop (Antares VIZIO, Work in Progress)
forum thread - http://forum.unity3d.com/viewtopic.php?t=31132

Nimbus Volumetric Clouds
forum thread - http://forum.unity3d.com/viewtopic.php?t=51972

RapidUnity Vehicle Editor Resource Pack
forum thread - http://forum.unity3d.com/viewtopic.php?t=56257

Ocean 3D - Ocean Simulation
website: http://www.nikko3d.com/index.php/downloads-2/

ShaderFusion - Node-Based Shader Editor (Requires Unity3)
forum thread - http://forum.unity3d.com/viewtopic.php?t=60611

Strumpy Shader Editor - Node-Based Shader Editor
forum thread - http://forum.unity3d.com/threads/56180-Strumpy-Shader-Editor-Beta-3.1-released-(performance-usability-upgrade)

Overview of extensions on UnifyWiki
http://www.unifycommunity.com/wiki/index.php?title=Extensions

LightUp (purchase required)
Extension of Google Sketchup (Lighting Solution) which works nicely for exporting lightmaps to Unity
http://www.light-up.co.uk/

Community Project - GTAIV Vehicle Replica (Pledge of > USD 50 required)
forum thread - http://forum.unity3d.com/viewtopic.php?t=50493

Stereoscopic Solutions
3D Anaglyph System (purchase required)
Stereoskopix3D

Plugin for Kinect's Primesense Camera
forum thread - http://forum.unity3d.com/threads/67982-Kinect-plugin

Unity Web Suite - tutorials and examples in C# to create online content
forum thread - http://forum.unity3d.com/threads/17674-Unity-Web-Suite-(Examples-Tutorials)
website - http://www.nickbreslin.com/2009/04/phphandler-20-released/

Tools for Visual Programming:
Antares Universe - Vizio (forum thread)
cost (as of 05/22/11) Euro 142.50
Visual programming tool similar to the approaches of Quest3D and Virtools.

Playmaker by Hutong Games
cost (as of 05/22/11) Euro 95.00
Visual State Machine Editor (website and forum thread)

uScript by Detox Studios (Beta Version)
website and forum thread
Visual Scripting Tool based on UDK's Kismet

Posted via email from Troy's posterous

Resolving nvcc & gcc conflict ion for Theano

2011-09-01T21:47:00.001+08:00

There is a compatibility issue affecting some Ubuntu 9.10 users, and probably anyone using CUDA 2.3 with gcc-4.4. Symptom: errors about “__sync_fetch_and_add” being undefined. Solution 1: make gcc-4.3 the default gcc (http://pascalg.wordpress.com/2010/01/14/cuda-on-ubuntu-9-10linux-mint-helena/) Solution 2: make another gcc (e.g. gcc-4.3) the default just for nvcc. Do this by making a directory (e.g. $HOME/.theano/nvcc-bindir) and installing two symlinks in it: one called gcc pointing to gcc-4.3 (or lower) and one called g++ pointing to g++-4.3 (or lower). Then add compiler_bindir = /path/to/nvcc-bindir to the [nvcc] section of your .theanorc (libdoc_config).

.theanorc file should be under the folder $HOME. If there is no such file, then create a new one.

From: http://deeplearning.net/software/theano/tutorial/using_gpu.html

Posted via email from Troy's posterous

Papers on learning sparse Gaussian Precision Matrix

2011-08-24T21:54:00.001+08:00

The basic idea of learning a sparse full covariance for the Gaussian using Graphical models was proposed in the paper "Covaraince Selection".

In the paper "Sparse Gaussian Graphical Models for Speech Recognition", the authors adopted the sparse full covariance learning to speech recognition.

In the third paper, "Projected Subgradient Methods for Learning Sparse Gaussians", a new approach for estimating the covaraince is proposed.

covariance selection.pdf Download this file

Sparse Gaussian Graphical Models for speech recognition_is2007.pdf Download this file

projected subgradient methods for learning sparse gaussians.pdf Download this file

Posted via email from Troy's posterous

LangBrain Website

2011-08-24T09:40:00.001+08:00

http://www.ruf.rice.edu/~lngbrain/main.htm

"If the human mind were simple enough for us to understand,
we would be too simple-minded to understand it."

Anonymous

The ability of humans to speak and to understand speech requires an enormous amount of brain resources. These resources have to manage information about many thousands of words and many syntactic constructions and their interconnections, not just to one another but to meanings and to the structures that allow us to recognize the sounds of speech and to move the muscles of our mouths to produce speech. This complex combination of brain structures can be called the brain's linguistic system. It allows a person not only to talk and to understand speech but also to read and write. It also gives us the power to think as well as the power to acquire new knowledge and abilities and to learn how to speak in the first place. The Langbrain website is about this system.

Posted via email from Troy's posterous

Graphical Gaussian Models for Genome Data

2011-08-23T21:48:00.001+08:00

From: http://strimmerlab.org/notes/ggm.html

Synopsis:

For software to efficiently identify GGM networks from data visit theGeneNet page.

A simple method for inferring the network of (linear) dependencies among a set of variables is to compute all pairwise correlations and subsequently to draw the corresponding graph (for some specified threshold). While popular and often used on many types of genomic data (e.g. gene expression, metabolite concentrations etc.)the naive correlation approach does not allow to infer the dependency network. Instead, graphical Gaussians models (GGMs) should be used. These allow to correctly identify direct influences, have close connections with causal graphical models, are straightforward to interpret, and yet are essentially as easy to compute as naive correlation models. This page lists pointers to learning GGMs from data, including procedures suitable for "small n, large p" data sets (category iii).

Introduction:

Graphical Gaussian Models (GGMs), also known as "covariance selection" or " concentration graph" models, have recently become a popular tool to study gene association networks. The key idea behind GGMs is to use partial correlations as a measure of independence of any two genes. This makes it straightforward to distinguish direct from indirect interactions. Note that partial correlations are related to the inverse of the correlation matrix. Also note that in GGMs missing edges indicate conditional independence.

A related but completely different concept are the so-called gene relevance networks which are based on the "covariance graph" model. In the latter interactions are defined through standard correlation coefficients so that missing edges denote marginal independence only.

There is a simple reason why GGMs should be preferred over relevance networks for identification of gene networks: the correlation coefficient is weak criterion for measuring dependence, as marginally, i.e. directly and indirectly, more or less all genes will be correlated. This implies that zero correlation is in fact a strong indicator for independence, i.e. the case of no edge in a network - but this is of course not what one usually wants to find out by building a relevance network... On the other hand, partial correlation coefficients do provide a strong measure of dependence and, correspondingly, offer only a weak criterion of independence (as most partial correlations coefficients usually vanish).

The best starting place to learn about GGMs is the classic paper that introduced this concept in the early 1970s. (A.P. Dempster. 1972. Covariance Selection. Biometrics 28:157-175). Further details can be found in the GGM books by J. Whittaker (1990) and by D. Edwards (1995).

Application of GGMs to Genomic Data:

Application of GGMs to genomic data is quite challenging, as the number of genes (p) is usually much larger than the number of available samples (n), and classical GGM theory is not valid in a small sample setting. With this page I'd like to provide a commented list of some recent work dealing with GGM gene expression analysis (there are only very few so far). In my understanding, all of these paper fit in one of three categories:

analysis with classic GGM theory,
using limited order partial correlations, and
application of regularized GGMs.

For small n, large p data it seems that methods from section iii. are most suited (see below for references and software).

I. Classic GGM Analysis:

The following papers simply apply classical GGM theory (i.e. with not further modification) to analyze gene expression data. It turns out that such an analysis is necessarily restricted to very small numbers of genes or gene clusters as to satisfy n > p.

P. J. Waddell and H. Kishino. 2000. Correspondence analysis of genes and tissue types and finding genetics links from microarray data. Genome Informatics 11:83-95
P. J. Waddell and H. Kishino. 2000. Cluster inferences methods and graphical models evaluated on NCI60 microarray gene expression data. Genome Informatics 11:129--140
H. Toh and K. Horimoto. 2002. Inference of a genetic network by a combined approach of cluster analysis and graphical Gaussian modeling. Bioinformatics 18:287--297
H. Toh and K. Horimoto. 2002. System for automatically inferring a genetic network from expression profiles. J. Biol. Physics 28:449--464
X. Wu, Y. Ye and K. R. Subramanian. 2003. Interactive analysis of gene interactions using graphical Gaussian model. ACM SIGKDD Workshop on Data Mining in Bioinformatics 3:63-69

II. Limited Order Partial Correlations:

One way to circumvent the problem of computing full partial correlation coefficients when the sample size is small compared to the number of genes is to use partial correlation coefficients of limited order. This results in something inbetween a full GGM model (with correlation conditioned on all p-2 remaining genes) and a relevance network model (with unconditioned correlation). This is the strategy employed in the following papers:

A. de la Fuente, N. Bing, I. Hoeschele, and P. Mendes. 2004. Discovery of meaningful associations in genomic data using partial correlation coefficients. Bioinformatics 20:3565-3574.
A. Wille, P. Zimmermann et al. 2004. Sparse graphical Gaussian modeling of the isoprenoid gene network in Arabidopsis thaliana. Genome Biology 5:R92
P. M. Magwene and J. Kim. 2004. Estimating genomic coexpression networks using first-order conditional independence. Genome Biology 5:R100
A. Wille and P. Bühlmann. 2006. Low-order conditional independence graphs for inferring genetic networks. Statist. Appl. Genet. Mol. Biol. 4: 32.
R. Castelo and A. Roverato. A robust procedure for Gaussian graphical model search from microarray data with p larger than n. Preprint.

III. Regularized GGMs:

Another possibility (and in my opinion the statistically most sound way) to marry GGMs with small sample modeling is to introduce regularization and moderation. This essentially boils down to finding suitable estimates for the covariance matrix and its inverse when n < p. This can either be done in a full Bayesian manner, or in an empirical Bayes way via variance reduction, shrinkage estimates etc. Once regularized estimates of partial correlation are available then heuristic searches can subsequently to be employed to find an optimal graphical model (or set of models).

Outside a genomic context using regularized GGMs was first proposed by F. Wong, C.K. Carter, and R. Kohn. (2003. Efficient estimation of covariance selection models. Biometrika 90:809-830). For gene expression data this strategy is pursued in the following papers:

A. Dobra, C. Hans, B. Jones, J.R. Nevins, and M. West. 2004. Sparse graphical models for exploring gene expression data. J. Multiv. Analysis 90:196-212.

See the web page of M. West for various other related articles.
In these papers a regularized estimate of the correlation matrix is obtained, either by Stein-type shrinkage (3) or by bootstrap variance reduction (2). This estimate is subsequently employed for computing partial correlation. Network selection is based on false discovery rate multiple testing. This method is implemented in GeneNet.
J. Schäfer and K. Strimmer. 2005. An empirical Bayes approach to inferring large-scale gene association networks. Bioinformatics 21: 754-764.
J. Schäfer and K. Strimmer. 2005. A shrinkage approach to large-scale covariance matrix estimation and implications for functional genomics. Statist. Appl. Genet. Mol. Biol. 4: 32.
J. Schäfer and K. Strimmer. 2005. Learning large-scale graphical Gaussian models from genomic data. In: J. F. Mendes. (Ed.). Proceedings of "Science of Complex Networks: from Biology to the Internet and WWW" (CNET 2004), Aveiro, PT, August 2004. (Publisher: The American Institute of Physics).
N. Mainshausen and P. Bühlmann 2006. High-dimensional graphs and variable selection with the lasso. Annals of Statistics 34 (3)

This approach uses lasso regression to induce sparsity on a node level among the partial correlations.
H. Li and J. Gui. 2006. Gradient directed regularization for sparse Gaussian concentration graphs, with applications to inference of genetic networks. Biostatistics 2006 7:302-317.

These authors regularize the concentration matrix rather than the covariance matrix.

Posted via email from Troy's posterous

[Math] The Matrix Cookbook

2011-08-22T15:19:00.001+08:00

The Matrix Cookbook.pdf Download this file

A good collection of matrix algebra!

Posted via email from Troy's posterous

HTK lattice format

2011-08-09T14:53:00.001+08:00

The lattices generated by HVite have the following general form

VERSION=1.0

UTTERANCE=testf1.mfc
lmname=wdnet
lmscale=20.00 wdpenalty=-30.00

vocab=dict
N=31
L=56

I=0
t=0.00
I=1

t=0.36
I=2
t=0.75

I=3
t=0.81
... etc

I=30
t=2.48
J=0

S=0
E=1
W=SILENCE

J=1
S=1
E=2

W=FOUR
... etc
J=55

S=29
E=30
W=SILENCE

v=0 a=-3239.01 l=0.00
v=0 a=-3820.77 l=0.00

v=0 a=-246.99 l=-1.20

The first 5 lines comprise a header which records names of the files used to generated the lattice along with the settings of the language model scale and penalty factors.

Each node in the lattice represents a point in time measured in seconds and each arc represents a word spanning the segment of the input starting at the time of its start node and ending at the time of its end node. For each such span, v gives the number of pronunciation used, a gives the acoustic score and l gives the language model score.

The language model scores in output lattices do not include the scale factors and penalties. There are removed so that the lattice can be used as a constrained network for subsequent recognizer testing.

When using HVite normally, the word level network file is specified using the -w option. When the -w option is included but no file name is included, HVite constructs the name of a lattice file from the name of the test file and inputs that. Hence, a new recognition network is created for each input file and recognition is very fast. This is an efficient way, for example, of experimentally determining optimum values for the language model scale and penalty factors.

Posted via email from Troy's posterous

A hierarchical context dependent neural network architecture for improved phone recognition

2011-08-04T20:35:00.001+08:00

a hierarchical context dependent neural network architecture for improved phone recognition.pdf Download this file

This paper explored the CD phone recognition on TIMIT using hybrid NN/HMM system.

1) Using two nets in tandem, one for CI posteriors and the other modeling the contexts from the CI posteriors;

2) Directly train a NN for CD state posteriors, too many outputs and not robust;

3) The first net is to give bottleneck features and then use another net on top of it.

The best results on TIMIT is 21.24% on core test set, there is less than 1% difference from the DBN based monophone recognition, thus rendering the context gain not significant. And also using DBN, the current best results is around 19%.

Posted via email from Troy's posterous

Dirichlet mixture models of neural net posteriors for hmm based speech recognition

2011-08-04T20:34:00.001+08:00

dirichlet mixture models of neural net posteriors for hmm based speech recognition.pdf Download this file

In this paper, the authors propose to using Dirichlet Mixture models instead of Gaussian Mixture models for the hybrid NN/HMM system. In the conventional NN/HMM system, the NN's posteriors are Gaussianized to feed into the HMM framework. However, as the posterior probabilities are lying on probability simplex and their distribution could be modeled by Dirichlet distributions. Thus Dirichlet Mixture models would be more preferable to Gaussian Mixture model.

However, the final system performance, although better than the GMM based system, are still far for the state-of-the-art performance.

Posted via email from Troy's posterous

Automatic speech recognition using hidden conditional neural fields

2011-08-04T20:25:00.001+08:00

automatic speech recongition using hidden conditional neural fields.pdf Download this file

The concept is not that new, is similar to Conditional Neural Network or Neural Conditional Random Fields. And the results on TIMIT is far from DBN based systems.

Posted via email from Troy's posterous

Slides for Introduction to Restricted Boltzmann Machine

2011-08-03T10:14:00.001+08:00

Frean1.pdf Download this file

Why RBMs could be stacked and learning them in a layer-wise greedy fashion?

Page 50-57

Posted via email from Troy's posterous

[HTK] Compile HTK on 64bit machine

2011-07-13T16:40:00.001+08:00

When compile HTK on 64bit machines, by default it will still try to compile the tools to 32bit. Then there will be an error after issue the command "make all":

/usr/include/gnu/stubs.h:7:27: fatal error: gnu/stubs-32.h: No such file or directory

To solve this problem, what we need is to install the multilib package for the gcc and g++. On Ubuntu, install following packages:

g++-multilib

gcc-multilib

and redo "make all".

Posted via email from Troy's posterous

Ubuntu 9.04 nVidia Driver Screen Resolution Problem

2011-07-13T15:55:00.001+08:00

Many people running Ubuntu 9.04 are having trouble with the proprietary nVidia driver (nvidia-graphics-driver-180 in my case) including getting it to go to high resolutions that fit the native resolution of widescreen monitors. I had the same problem with an nVidia GeForce 6150 LE and Dell UltraSharp 2407WFPHC monitor.

I was able to get all the resolutions, including 1920×1200, for my monitor as well as have the nVidia driver recognize the monitor as a 2407WFPHC, by doing the following:

(If you can’t see your screen at all after enabling the nVidia driver, first read the companion post, Ubuntu 9.04 Screen Resolution/Monitor Out of Range (nVidia Driver 180).)

Open a terminal window
Go to the X11 directory (cd /etc/X11)
Make a backup of the current xorg.conf (e.g., sudo cp xorg.conf xorg.conf.backup)
Run nvidia-xconfig with root permission (sudo nvidia-xconfig). If you get a parsing error, delete xorg.conf so nvidia-xconfig can create a fresh one.
Open xorg.conf with your favorite editor (e.g. sudo vim xorg.conf)
You’ll see a lot of extra settings now
Look for Section “Monitor”. Mine defaulted to the following settings:
    Identifier “Monitor0″
    VendorName “Unknown”
    ModelName “Unknown”
    HorizSync 28.0 – 33.0
    VertRefresh 43.0 – 72.0
    Option “DPMS”
Change the HorizSync and VertRefresh values to the correct ones for your particular monitor. For my 2407WFPHC, I put the following:
HorizSync 30.0 – 83.0
VertRefresh 56.0 – 76.0
Save the xorg.conf file
Log out and restart the X server (at the login screen, select Menu, then Restart X server)
Log in and run the NVIDIA X Server Settings tool. You should now have a whole bunch of resolutions from which to choose. I selected 1920×1200.

The reason that this works is that the nVidia driver needs to know the frequency ranges for your monitor in order to know what resolutions are safe to use. Setting the HorizSync and VertRefresh in xorg.conf provides this necessary information.

From:

http://turbulentsky.com/ubuntu-904-nvidia-driver-screen-resolution-problem.html

Posted via email from Troy's posterous

[Matlab] staring error on Ubuntu - libc.so.6 not found

2011-06-08T11:30:00.001+08:00

MATLAB with Ubuntu 11.04

Under the new Ubuntu 11.04 (Natty Narwhal), when you try and run MATLAB you'll get the following error

/matlab/bin/util/oscheck.sh: 605: /lib64/libc.so.6: not found

To resolve this problem, go to a terminal window (Alt + F2 from anywhere) and type

For 64 bit:

sudo ln -s /lib64/x86_64-linux-gnu/libc-2.13.so /lib64/libc.so.6

For 32 bit:

sudo ln -s /lib/i386-linux-gnu/libc-2.13.so /lib/libc.so.6

This should restore the missing library that was uninstalled during the Ubuntu update process.

From: http://morganbye.net/blog/2011/05/matlab-ubuntu-1104

Posted via email from Troy's posterous

[CUDA] Solving the problem of not supporting gcc-4.5 and above

2011-06-07T20:43:00.001+08:00

Here is what I usually do:
- in parallel to the current gcc version (currently 4.5.2 on Arch Linux), install an older one. Right now I have 4.3.4 and 4.4.4, available as "gcc-4.3" and "gcc-4.4".
- for each of these old compilers, create a folder, say /opt/gcc43 and /opt/gcc44, which contain symlinks to /usr/bin/g{cc,++}-4.{3,4} that are named "gcc" and "g++" without version numbers
- run nvcc with the --compiler-bindir option: "nvcc --compiler-bindir /opt/gcc44 $CUFLAGS ..."

From: http://forums.nvidia.com/index.php?showtopic=194872

Posted via email from Troy's posterous

Supervised Clustering

2011-06-02T23:35:00.001+08:00

See Also:

Launch in a standalone WM Player
Switch to Windows Media Player

Download slides: nips2010_zadeh_sc_01.pdf (82.6 KB)

Download article: nips2010_0427.pdf (93.9 KB)

View slides

Streaming Video Help

Windows Media Player Firefox Plugin - Download

via videolectures.net

Doing clustering with teachers.

Posted via email from Troy's posterous

[ASR] Langue model

2011-06-01T22:08:00.001+08:00

HVite can only use bigram LM after converted to lattice format.

When generate the bigram model, it should not be count based LM,i.e. not use "-x" in LBuild.

Incorrect:

LBuild -T 1 -C lfof.cfg -c 2 0 -x -n 2 timitphns.0/wmap bg0c timitphns.0/gram.0

Correct:

LBuild -T 1 -C lfof.cfg -c 2 0 -n 2 timitphns.0/wmap bg0c timitphns.0/gram.0

Otherwise, after converted to lattice,

HBuild -n bigram -s "<s>" "</s>" word61.lst bigram_lat

HVite cannot work correctly. For my experiments, using the incorrect LM, the recognized results are all:

Posted via email from Troy's posterous

[Tools] HVite for alignment

2011-05-31T20:38:00.001+08:00

When using HVite to do the alignment, we need the option '-a'. One thing to be careful is, Don't use '-w' for alignment!!!!

Otherwise, it will load the network from '-w' not '-a', and then recognition is carried out instead of alignment, even though you put '-a' there!

Incorrect:

HVite -A -D -V -T 1 -X rec -l '*' -o SW -a -m -f -X lab -y lab -H ${exppath}/hmm648/MMF -C ${exppath}/cfgs/hvite.cfg -t 250.0 -s 4.0 -p -10.0 -w ${libpath}/lms/bigram_lat -I ${libpath}/mlabs/test_mono61_label.mlf -i ${exppath}/hmm648/test_mono61_alnlab.mlf -S ${libpath}/flists/testfull_mfcc.scp ${libpath}/dicts/mono61.dct ${libpath}/mlists/mono61.lst

Correct:

HVite -A -D -V -T 1 -X rec -l '*' -o SW -a -m -f -X lab -y lab -H ${exppath}/hmm648/MMF -C ${exppath}/cfgs/hvite.cfg -t 250.0 -I ${libpath}/mlabs/test_mono61_label.mlf -i ${exppath}/hmm648/test_mono61_alnlab.mlf -S ${libpath}/flists/testfull_mfcc.scp ${libpath}/dicts/mono61.dct ${libpath}/mlists/mono61.lst

Posted via email from Troy's posterous

[Tools] Compiling Matlab .m Files

2011-05-25T16:24:00.001+08:00

From: http://pages.stern.nyu.edu/~nwhite/scrc/CompilingMatlab.htm

Compiling Matlab .m Files

Norman White, November 2006

Matlab allows users to compile their matlab programs into executable images that can be run anywhere without license restrictions. This is particularly useful in a grid computing environment like Stern's, where a user may want to run many jobs at the same time, but be limited by matlab and toolbox license constraints. Until recently, the matlab compiler was difficult to use, but in a recent release, Matlab 73 (2006b), it has been significantly improved. The information below will show you how to convert your matlab .m files into matlab executables.

1) ssh to rnd.

That is where the matlab compiler lives. Since we have many linux nodes and only one Sun Solaris node, it makes sense to generate an intel linux executable. We only have a single user license, but that should be enough, since you only have to compile once. Note, that every time you compile, matlab will lock the compiler license for 30 minutes. You can continue to use it, but a new user will have to wait for 30 minutes after you have finished.

a) Edit your matlab.m file and turn it into a function. This can be done by putting

function myfun

statement as the first line.

b) Type (You only need to do this once)

mbuild -setup

to initialize the matlab compile environment.

c) If you have any matlab.m

files that are called by your main matlab file, make sure they are on the path, or in the directory

you are compiling matlab in (For instance, if you have some matlab files in a directory call mylib, execute the following command before compiling).

Export PATH="myfile:$PATH".

Note that matlab is going to compile ALL of the matlab code that is needed to run your program, including matlab toolboxes.

d) Compile the program by typing

mcc -m mymatjob.m

This should grind along for a few minutes and hopefully complete with no errors.

When it is done, there will be an executable file named myfun in your directory. It

can now be run on any node that has access to the matlab libraries. To make sure

your job runs on the appropriate nodes, add a

#$ -l matcomp=1

line to your shell script that will run the program (see below)

2) Create a shell script that will setup the libraries needed to run and will invoke your program.

Create a shell script that looks like this.

#! /bin/sh

# Name: myfun.sh

# Run myfun compiled program

# Set up matlab LD_LIBRARY_PATH

LD_LIBRARY_PATH="/usr/local/matlab73/bin/glnx86":"/usr/local/matlab73/sys/os/glnx86"

# Then run your program

./myfun

and then save it as myfun.sh. After saving it, make sure it is readable and executable by others by typing

chmod o+r myfun.sh

3) Create the Sun Grid Engine shell script to run the program. Here is an example:

#! /bin/sh

# Name: Runmyfun.sh

#$ -N Runmyfun.sh

#$ -l matcomp=1

#$ -l h_cpu=10:00:00 (Optional, put your cpu requirements here)

#$ -l h_vmem=1000m (This is optional, put your memory requirements here).

# Now run the program

/bin/sh myfun.sh

Save the program as Runmyfun.sh, and then

chmod o+rx Runmyfun.sh

4) Submit the job..

qsub -t 1-300 Runmyfun.sh

This will run the job on as many nodes as are available, but not take up ANY matlab licenses (not matlab or toolboxes). The toolbox .m files will already be compiled into your code. For instance, we only have two optimization toolbox license. So you could only run your code on two systems at a time, instead of 10.

This same code can also be run on almost any grid computing environment, just by moving the executable file and a few other files. Again, no need for a matlab license.

Have fun�

Posted via email from Troy's posterous

[ASR] A complete Matlab script for TIMIT phoneme recognition

2011-05-24T12:33:00.001+08:00

htk.m Download this file

% HTK-TIMIT Phone Recognition % Tae-Jin Yoon % tyoon@uvic.ca % % The script is taken from somewhere on the web, which is now defunct. % I modified the script to run on my Windows Vista computer. % So, to run this you need to have a TIMIT corpus and need to change the % configuration that are specific to my computer. % I resampled the sampling rate of the TIMIT corpus to 16kHz. htkdir = ['c:\MATLAB\R2008a\work\htk\']; homedir = ['c:\work\timit\htk\']; traindir = ['c:\work\timit\TIMIT\train\']; testdir = ['c:\work\timit\TIMIT\test\']; eval(['!']); eval(['!mkdir label']); eval(['!mkdir mfcc']); eval(['!mkdir model']); for i=0:23 eval(['!mkdir model\hmm', num2str(i)]) end %% Make a gram file and rune HParse disp('make a gram file'); fid = fopen('gram', 'w'); fprintf(fid, '%s\n', ['$beginend = h#;']); fprintf(fid, '%s', ['$phone = bcl | b | dcl | d | gcl | g | pcl | p | tcl | t | kcl | k | ']); fprintf(fid, '%s', ['dx | q | jh | ch | s | sh | z | zh | f | th | v | dh | m | n | ng | em | en | eng | ']); fprintf(fid, '%s', ['nx | l | r | w | y | hh | hv | el | ']); fprintf(fid, '%s', ['iy | ih | eh | ey | ae | aa | aw | ay | ah | ao | oy | ow | ']); fprintf(fid, '%s\n', ['uh | uw | ux | er | ax | ix | axr | ax-h | pau | epi;']); fprintf(fid, '%s\n', ['($beginend <$phone> $beginend)']); fclose(fid); %input('Press enter to continue'); eval(['!c:\MATLAB\R2008a\work\htk\HParse.exe -T 1 gram wdnet']); disp('Make a monophones0 file. The file contains phone symobls'); fid = fopen('monophones0', 'w'); fprintf(fid, '%s\n', 'b'); fprintf(fid, '%s\n', 'd'); fprintf(fid, '%s\n', 'g'); fprintf(fid, '%s\n', 'p'); fprintf(fid, '%s\n', 't'); fprintf(fid, '%s\n', 'k'); fprintf(fid, '%s\n', 'dx'); fprintf(fid, '%s\n', 'q'); fprintf(fid, '%s\n', 'jh'); fprintf(fid, '%s\n', 'ch'); fprintf(fid, '%s\n', 's'); fprintf(fid, '%s\n', 'sh'); fprintf(fid, '%s\n', 'z'); fprintf(fid, '%s\n', 'zh'); fprintf(fid, '%s\n', 'f'); fprintf(fid, '%s\n', 'th'); fprintf(fid, '%s\n', 'v'); fprintf(fid, '%s\n', 'dh'); fprintf(fid, '%s\n', 'm'); fprintf(fid, '%s\n', 'n'); fprintf(fid, '%s\n', 'ng'); fprintf(fid, '%s\n', 'em'); fprintf(fid, '%s\n', 'en'); fprintf(fid, '%s\n', 'eng'); fprintf(fid, '%s\n', 'nx'); fprintf(fid, '%s\n', 'l'); fprintf(fid, '%s\n', 'r'); fprintf(fid, '%s\n', 'w'); fprintf(fid, '%s\n', 'y'); fprintf(fid, '%s\n', 'hh'); fprintf(fid, '%s\n', 'hv'); fprintf(fid, '%s\n', 'el'); fprintf(fid, '%s\n', 'iy'); fprintf(fid, '%s\n', 'ih'); fprintf(fid, '%s\n', 'eh'); fprintf(fid, '%s\n', 'ey'); fprintf(fid, '%s\n', 'ae'); fprintf(fid, '%s\n', 'aa'); fprintf(fid, '%s\n', 'aw'); fprintf(fid, '%s\n', 'ay'); fprintf(fid, '%s\n', 'ah'); fprintf(fid, '%s\n', 'ao'); fprintf(fid, '%s\n', 'oy'); fprintf(fid, '%s\n', 'ow'); fprintf(fid, '%s\n', 'uh'); fprintf(fid, '%s\n', 'uw'); fprintf(fid, '%s\n', 'ux'); fprintf(fid, '%s\n', 'er'); fprintf(fid, '%s\n', 'ax'); fprintf(fid, '%s\n', 'ix'); fprintf(fid, '%s\n', 'axr'); fprintf(fid, '%s\n', 'ax-h'); fprintf(fid, '%s\n', 'bcl'); fprintf(fid, '%s\n', 'dcl'); fprintf(fid, '%s\n', 'gcl'); fprintf(fid, '%s\n', 'pcl'); fprintf(fid, '%s\n', 'tcl'); fprintf(fid, '%s\n', 'kcl'); fprintf(fid, '%s\n', 'pau'); fprintf(fid, '%s\n', 'epi'); fprintf(fid, '%s\n', 'h#'); fclose(fid) %% PREPARING TRAINING DATA %% %%%%%%%%%%%%%%%%%%%%%%%%%%%%% fid1 = fopen('codetr.scp', 'w'); fid2 = fopen('train.scp', 'w'); %fid3 = fopen('fname.praat', 'w'); for n0=1:8 D = dir([traindir, 'DR', num2str(n0)]); for n1=3:size(D,1) D2 = dir([traindir, 'DR', num2str(n0),'\',D(n1).name,'\*.wav']); D3 = dir([traindir, 'DR', num2str(n0),'\',D(n1).name, '\*.phn']); for n2 = 1:size(D2,1) filename = [traindir, 'DR', num2str(n0), '\', D(n1).name, '\' D2(n2).name]; % fprintf(fid3, '%s\n', ['Read from file... ', filename]); % fprintf(fid3, '%s\n', ['Write to WAV file... ', filename]); % fprintf(fid3, '%s\n', 'Remove'); handdefname = [traindir, 'DR', num2str(n0), '\',D(n1).name, '\' D3(n2).name]; newfname=D2(n2).name; newfname=[newfname(1:end-4) '_tr.mfc']; mfcfname = [homedir, 'mfcc\', 'DR', num2str(n0), '_', D(n1).name, '_', newfname]; fprintf(fid1, '%s\n', [filename, ' ', mfcfname]); fprintf(fid2, '%s\n', mfcfname); newlname = D3(n2).name; newlname = [newlname(1:end-4) '_tr.lab']; labfname = [homedir, 'label\', 'DR', num2str(n0), '_', D(n1).name, '_', newlname]; eval(['!copy ', handdefname, ' ', labfname]); end end end fclose(fid1); fclose(fid2); % fclose(fid3); disp('The original TIMIT wav files need to be converted to MSWAVE file format'); disp('Make a praat script'); input('Press enter to continue'); %% PREPARE TIMIT configuration file & run HCopy fid = fopen('configTIMIT', 'w'); fprintf(fid, '%s\d', ['SOURCEKIND = WAVEFORM']); fprintf(fid, '%s\d', ['SOURCEFORMAT = WAV']); fprintf(fid, '%s\d', ['SOURCERATE = 625']); fprintf(fid, '%s\d', ['TARGETKIND = MFCC_0_D_A']); fprintf(fid, '%s\d', ['TARGETRATE = 100000.0']); fprintf(fid, '%s\d', ['WINDOWSIZE = 250000.0']); fprintf(fid, '%s\d', ['USEHAMMING = T']); fprintf(fid, '%s\d', ['PREEMCOEF = 0.97']); fprintf(fid, '%s\d', ['NUMCHANS = 20']); fprintf(fid, '%s\d', ['CEPLIFTER = 22']); fprintf(fid, '%s\d', ['NUMCEPS = 12']); fclose(fid); eval(['!c:\MATLAB\R2008a\work\htk\HCopy.exe -T 1 -C configTIMIT -S codetr.scp']); %% PREPARING TESTING DATA %% %%%%%%%%%%%%%%%%%%%%%%%%%%%% fid1 = fopen('codete.scp', 'w'); fid2 = fopen('test.scp', 'w'); %fid3 = fopen('test.praat', 'w'); for n0 = 1:8 D = dir([testdir, 'DR', num2str(n0)]); for n1 = 3:size(D, 1) D2 = dir([testdir, 'DR', num2str(n0), '\', D(n1).name, '\*.wav']); D3 = dir([testdir, 'DR', num2str(n0), '\', D(n1).name, '\*.phn']); %#%disp(D2(n1).name); for n2 = 1:size(D2, 1) filename=[testdir, 'DR', num2str(n0), '\', D(n1).name, '\' D2(n2).name]; % fprintf(fid3, '%s\n', ['Read from file... ', filename]); % fprintf(fid3, '%s\n', ['Write to WAV file... ', filename]); % fprintf(fid3, '%s\n', 'Remove'); handdefname = [testdir, 'DR', num2str(n0), '\', D(n1).name, '\' D3(n2).name]; newfname = D2(n2).name; newfname=[newfname(1:end-4) '_te.mfc']; mfcfname = [homedir, 'mfcc\', 'DR', num2str(n0), '_', D(n1).name, '_', newfname]; fprintf(fid1, '%s\n', [filename, ' ', mfcfname]); fprintf(fid2, '%s\n', mfcfname); newlname = D3(n2).name; newlname = [newlname(1:end-4) '_te.lab']; labfname = [homedir, 'label\', 'DR', num2str(n0), '_', D(n1).name, '_', newlname]; eval(['!copy ', handdefname, ' ', labfname]); end end end fclose(fid1); fclose(fid2); % fclose(fid3); eval(['!c:\MATLAB\R2008a\work\htk\HCopy.exe -T 1 -C configTIMIT -S codete.scp']); %% PREPARE configuration file fid = fopen('config', 'w'); fprintf(fid, '%s\d', ['TARGETKIND = MFCC_0_D_A']); fprintf(fid, '%s\d', ['TARGETRATE = 100000.0']); fprintf(fid, '%s\d', ['WINDOWSIZE = 250000.0']); fprintf(fid, '%s\d', ['USEHAMMING = T']); fprintf(fid, '%s\d', ['PREEMCOEF = 0.97']); fprintf(fid, '%s\d', ['NUMCHANS = 20']); fprintf(fid, '%s\d', ['CEPLIFTER = 22']); fprintf(fid, '%s\d', ['NUMCEPS = 12']); fclose(fid) eval(['!c:\MATLAB\R2008a\work\htk\HCompV.exe -T 1 -C config -f 0.01 -m -S train.scp -M model/hmm0 proto']); disp('Make proto file'); fid = fopen('proto', 'w'); fprintf(fid, '%s\n', ['~o <VecSize> 39 <MFCC_0_D_A>']); fprintf(fid, '%s\n', ['~h "proto"']); fprintf(fid, '%s\n', ['<BeginHMM>']); fprintf(fid, '\t%s\n', ['<NumStates> 5']); fprintf(fid, '\t%s\n', ['<State> 2']); fprintf(fid, '\t\t%s\n', ['<Mean> 39']); fprintf(fid, '\t\t\t%s\n', ['0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0']); fprintf(fid, '\t\t%s\n', ['<Variance> 39']); fprintf(fid, '\t\t\t%s\n', ['1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1']); fprintf(fid, '\t%s\n', ['<State> 3']); fprintf(fid, '\t\t%s\n', ['<Mean> 39']); fprintf(fid, '\t\t\t%s\n', ['0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0']); fprintf(fid, '\t\t%s\n', ['<Variance> 39']); fprintf(fid, '\t\t\t%s\n', ['1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1']); fprintf(fid, '\t%s\n', ['<State> 4']); fprintf(fid, '\t\t%s\n', ['<Mean> 39']); fprintf(fid, '\t\t\t%s\n', ['0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0']); fprintf(fid, '\t\t%s\n', ['<Variance> 39']); fprintf(fid, '\t\t\t%s\n', ['1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1']); fprintf(fid, '\t%s\n', ['<TransP> 5']); fprintf(fid, '\t\t%s\n', ['0 1 0 0 0']); fprintf(fid, '\t\t%s\n', ['0 0.6 0.4 0 0']); fprintf(fid, '\t\t%s\n', ['0 0 0.6 0.4 0']); fprintf(fid, '\t\t%s\n', ['0 0 0 0.7 0.3']); fprintf(fid, '\t\t%s\n', ['0 0 0 0 0']); fprintf(fid, '%s\n', ['<EndHMM>']); fclose(fid) fid = fopen('model/hmm0/proto', 'r'); F = fread(fid); S = char(F'); SHMM = S( strfind(upper(S), '<BEGINHMM>') :end); S1st3 =S(1: strfind(S, '~h') -1); fclose(fid); fid = fopen('model/hmm0/vFloors', 'r'); F = fread(fid); SvFloors = char(F'); fclose(fid); fid = fopen('model/hmm0/macros', 'w'); fprintf(fid, S1st3); fprintf(fid, SvFloors); fclose(fid); fid1 = fopen('monophones0', 'r'); fid2 = fopen('model/hmm0/hmmdefs', 'w'); fid3 = fopen('dict', 'w'); fid4 = fopen('monophones1', 'w'); while 1 tline = fgetl(fid1); disp(tline); if ~ischar(tline) break; end fprintf(fid2, ['~h "', tline, '"\n']); fprintf(fid2, SHMM); fprintf(fid2, '\n'); fprintf(fid3, [tline, ' ', tline, '\n']); fprintf(fid4, [tline, '\n']); end fprintf(fid4,['!ENTER\n']); fprintf(fid4,['!EXIT\n']); fprintf(fid3,['!ENTER []\n']); fprintf(fid3,['!EXIT []\n']); fclose(fid1); fclose(fid2); fclose(fid3); fclose(fid4); % input('Press enter to continue'); fid1=fopen('phones0.mlf','w'); fid3=fopen('HLStatslist','w'); fprintf(fid1,'%s\n',['!#']); D=dir(['label/*tr.lab']); for n=1:size(D,1) fprintf(fid1,'%s\n',['"*/',D(n).name,'"']); fprintf(fid3,[D(n).name,'\n']); fid2=fopen(['label/',D(n).name],'r'); while 1 tline=fgetl(fid2); if ~ischar(tline) break; end; if (tline(1)=='#')|(tline(1)=='"') fprintf(fid1,'%s\n',tline); else Tmat=sscanf(tline,'%d %d %s'); Tstring=[char(Tmat(3:end))]'; fprintf(fid1,'%s\n',Tstring); end end fprintf(fid1,'%s\n','.'); fclose(fid2); end fprintf(fid1,'\n'); fclose(fid1); fclose(fid3); % input('Press enter to continue'); eval(['!c:\MATLAB\R2008a\work\htk\HLStats.exe -T 1 -b bigfn -o -I phones0.mlf monophones0 -S HLStatslist']); eval(['!c:\MATLAB\R2008a\work\htk\HBuild.exe -T 1 -n bigfn monophones1 outLatFile']); fid1 = fopen('testref.mlf','w'); fprintf(fid1, '%s\n', ['!#']); D = dir(['label/*te.lab']); for n = 1:size(D, 1) fprintf(fid1, '%s\n', ['"*\', D(n).name, '"']); fid2 = fopen(['label\', D(n).name], 'r') while 1 tline = fgetl(fid2); if ~ischar(tline) break; end if (tline(1) == '#') | (tline(1) == '"') fprintf(fid1, '%s\n', tline); else Tmat=sscanf(tline,'%d %d %s'); Tstring=[char(Tmat(3:end))]'; fprintf(fid1,'%s\n',Tstring); end end fprintf(fid1, '%s\n', '.'); fclose(fid2); end fprintf(fid1, '\n'); fclose(fid1); for i = 1:3 eval(['!c:\MATLAB\R2008a\work\htk\HERest.exe -T 1 -C config -I phones0.mlf -t 250.0 150.0 1000.0 -S train.scp -H model/hmm', num2str(i-1), '/macros -H model/hmm', num2str(i-1), '/hmmdefs -M model/hmm', num2str(i), ' monophones0']); end fid = fopen('sil.hed', 'w'); fprintf(fid, ['AT 2 4 0.2 {pau.transP}\n']); fprintf(fid, ['AT 4 2 0.2 {pau.transP}\n']); fprintf(fid, ['AT 2 4 0.2 {h#.transP}\n']); fprintf(fid, ['AT 4 2 0.2 {h#.transP}\n']); fclose(fid); eval(['!c:\MATLAB\R2008a\work\htk\HHEd.exe -T 1 -H model/hmm3/macros -H model/hmm3/hmmdefs -M model/hmm4 sil.hed monophones0']); for i = 5:7 eval(['!c:\MATLAB\R2008a\work\htk\HERest -T 1 -C config -I phones0.mlf -t 250.0 150.0 1000.0 -S train.scp -H model/hmm', num2str(i-1), '/macros -H model/hmm', num2str(i-1), '/hmmdefs -M model/hmm', num2str(i), ' monophones0']); end fid = fopen('MU2.hed', 'w'); fprintf(fid, ['MU 2 {*.state[2-4].mix}\n']); fclose(fid); eval(['!c:\MATLAB\R2008a\work\htk\HHEd.exe -T 1 -H model/hmm7/macros -H model/hmm7/hmmdefs -M model/hmm8 MU2.hed monophones0']); for i=9:11 eval(['!c:\MATLAB\R2008a\work\htk\HERest -T 1 -C config -I phones0.mlf -t 250.0 150.0 1000.0 -S train.scp -H model/hmm',num2str(i-1),'/macros -H model/hmm',num2str(i-1),'/hmmdefs -M model/hmm',num2str(i),' monophones0']); end fid=fopen('MU4.hed','w'); fprintf(fid,['MU 4 {*.state[2-4].mix}\n']); fclose(fid); %input('Press enter to continue'); eval(['!c:\MATLAB\R2008a\work\htk\HHEd.exe -T 1 -H model/hmm11/macros -H model/hmm11/hmmdefs -M model/hmm12 MU4.hed monophones0']); % input('Press enter to continue'); for i=13:15 eval(['!c:\MATLAB\R2008a\work\htk\HERest.exe -T 1 -C config -I phones0.mlf -t 250.0 150.0 1000.0 -S train.scp -H model/hmm',num2str(i-1),'/macros -H model/hmm',num2str(i-1),'/hmmdefs -M model/hmm',num2str(i),' monophones0']); end fid=fopen('MU8.hed','w'); fprintf(fid,['MU 8 {*.state[2-4].mix}\n']); fclose(fid); eval(['!c:\MATLAB\R2008a\work\htk\HHEd.exe -T 1 -H model/hmm15/macros -H model/hmm15/hmmdefs -M model/hmm16 MU8.hed monophones0']); for i=17:23 eval(['!c:\MATLAB\R2008a\work\htk\HERest.exe -T 1 -C config -I phones0.mlf -t 250.0 150.0 1000.0 -S train.scp -H model/hmm',num2str(i-1),'/macros -H model/hmm',num2str(i-1),'/hmmdefs -M model/hmm',num2str(i),' monophones0']); end eval(['!c:\MATLAB\R2008a\work\htk\HVite.exe -T 1 -H model/hmm23/macros -H model/hmm23/hmmdefs -S test.scp -i recout.mlf -w wdnet -p 0.0 -s 5.0 dict monophones0']); %input('Press enter to continue'); disp('With bigram language model:'); eval(['!c:\MATLAB\R2008a\work\htk\HVite.exe -T 1 -H model/hmm23/macros -H model/hmm23/hmmdefs -S test.scp -i recout_bigram.mlf -w outLatFile -p 0.0 -s 5.0 dict monophones0']); disp('Error occurs in recout.mlf'); disp('Need to change mfcc to label in recout.mlf'); disp('A hacky way using Vim:.,$,s/mfcc/label/g'); eval(['!c:\MATLAB\R2008a\work\htk\HResults.exe -T 1 -I testref.mlf monophones0 recout.mlf']); eval(['!c:\MATLAB\R2008a\work\htk\HResults.exe -T 1 -e n en -e aa ao -e ah ax-h -e ah ax -e ih ix -e l el -e sh zh -e uw ux -e er axr -e m em -e n nx -e ng eng -e hh hv -e pau pcl -e pau tcl -e pau kcl -e pau q -e pau bcl -e pau dcl -e pau gcl -e pau epi -e pau h# -I testref.mlf monophones0 recout.mlf']); %input('Press enter to continue'); % disp('With bigram language model:'); %eval(['!HResults -T 1 -I testref.mlf monophones0 recout_bigram.mlf']); % eval(['!c:\MATLAB\R2008a\work\htk\HResults -T 1 -e n en -e aa ao -e ah ax-h -e ah ax -e ih ix -e l el -e sh zh -e uw ux -e er axr -e m em -e n nx -e ng eng -e hh hv -e pau pcl -e pau tcl -e pau kcl -e pau q -e pau bcl -e pau dcl -e pau gcl -e pau epi -e pau h# -I testref.mlf monophones0 recout_bigram.mlf > results_bigram']);

Posted via email from Troy's posterous

[ASR] TIMIT information for phoneme recognition

2011-05-23T21:35:00.001+08:00

TIMIT.pdf Download this file

Posted via email from Troy's posterous

[NN] NN output activations with cross entropy error function

2011-05-16T13:57:00.001+08:00

NN_sigmoid_softmax.pdf Download this file

Binomial - sigmoid activation function;

Multinomial - softmax activation function.

The error functions are all cross entropy.

Posted via email from Troy's posterous

[Business] Business Plan Formating

2011-05-04T19:43:00.001+08:00

BUSINESS PLAN
FORMATTING

We like business plans that present a lot of information in as few words as possible. The following format, within 15-20 slides, is all that’s needed:

COMPANY PURPOSE

Define the company/business in a single declarative sentence.

PROBLEM

Describe the pain of the customer (or the customer’s customer).
Outline how the customer addresses the issue today.

SOLUTION

Demonstrate your company’s value proposition to make the customer’s life better.
Show where your product physically sits.
Provide use cases.

WHY NOW

Set-up the historical evolution of your category.
Define recent trends that make your solution possible.

MARKET SIZE

Identify/profile the customer you cater to.
Calculate the TAM (top down), SAM (bottoms up) and SOM.

COMPETITION

List competitors
List competitive advantages

PRODUCT

Product line-up (form factor, functionality, features, architecture, intellectual property).
Development roadmap.

BUSINESS MODEL

Revenue model
Pricing
Average account size and/or lifetime value
Sales & distribution model
Customer/pipeline list

TEAM

Founders & Management
Board of Directors/Board of Advisors

FINANCIALS

P&L
Balance sheet
Cash flow
Cap table
The deal

Posted via email from Troy's posterous

[Business] Business model

2011-04-28T19:28:00.001+08:00

What is your business model? Amazingly many people who start a business can't answer this question. In its most basic form, the business model is the diagram of how you will convert your product or service into money in your pocket.

Many entrepreneurs suffer from the Field of Dreams affliction. They are sure that their product or service is so fabulous that everyone will want it without taking into account how their potential customers buy things or even how they will finance the gap between the sale and the cash coming in the door.

To create your business model, you need to answer the following questions:

to whom are you going to sell your product?
what problems does your product or service solve for the customer?
how you are going to reach your customers?
how and when the customers will pay you?
how your products will get from concept stage into your customers' hands?
who are your competitors and how are you different?

In business speak, the business model needs the following components:

Value Proposition - this term is highly overused by every MBA and VC associate out there, but the concept is important. The value proposition is a description of the value that you bring to your customer; ie, what problem are you solving? In short, is your product faster, better, or cheaper?

Market Segment - who are you going to sell your product or service to. This is the type of customer that you will target. You need to determine why they are the right group who are willing to pay money to you.

Value Chain - another good b-school term. You need to determine where you are in the value chain. If you are making a widget that improves a driver's experience, but it can only be installed by the car manufacturer, then you have a place in the automotive value chain. You need to understand how you fit in this chain and what prices you can charge as a result of that place.

Revenue and Margin - How will you price the product and how will people pay you? License, fee upfront, payment over time. What are your costs? Will you be able to bring them down over time. If there is a big profit margin in your business, expect other company to try to enter the same business.

Competitive Strategy - who is your competition and how will you differentiate yourself from them. Are you better, faster or cheaper? Does your competition have the ability to improve in response to your entry into the market? Don't forget that your customer choosing to do nothing can be competition.

Many entrepreneurs think that the business model is the same as the business plan. Your business plan is actually how you plan to execute on your business model. Build the model first, find your strengths and weaknesses and write the plan on how you will reach your goal.

Article Source: http://EzineArticles.com/966230

Posted via email from Troy's posterous

[ML] Some learnings

2011-04-27T09:50:00.001+08:00

* Active Learning = supervised + oracle queries
* Semi-supervised Learning = supervised + unlabeled data
* Multitask & Transfer Learning = supervised + data from related task(s)
* Domain Adaptation = supervised + data from source domain(s)
* Multiview Learning = supervised + alternative data views
* Learning in a Bandit Setting = supervised + limited feedback
* Multiple-Instance Learning = supervised + relaxed label granularity
* Learning with Expert Knowledge = supervised + side information
* Weakly Labeled Data = supervised + free (possibly inaccurate) labels
* Human Computation and Crowd-sourcing = supervised + free/inexpensive
annotators
* Learning on a Budget = supervised + misc. cost constraints

Posted via email from Troy's posterous

[DBN] Deep Machine Learning - A New Frontier in Artificial Intelligence Research

2011-04-25T12:49:00.001+08:00

DML_Arel_2010.pdf Download this file

Posted via email from Troy's posterous

[DBN] Deep Learning and Its Applications to Signal and Information Processing

2011-04-20T10:37:00.001+08:00

DeepLearn-SPM2010.pdf Download this file

Dong Yu and Li Deng

exploratory DSP 2010

Posted via email from Troy's posterous

Pronunciation teaching - Internet Resources

2011-04-01T15:39:00.001+08:00

Check out this website I found at liceu.uab.es

http://liceu.uab.es/~joaquim/applied_linguistics/L2_phonetics/rec_Corr_Fon_in...

This page lists a quite large amount of resources for Computer Assisted Language Learning research.

Posted via email from Troy's posterous

[Latex] Simultaneous Equations in Latex

2011-03-31T16:44:00.001+08:00

Writing Simultaneous Equations in Latex using cases command, by default, it deems the whole group of equations as one equation and only gives one number for the whole set.

If we want to add numbering to each of the equations inside the group, it's a little difficult.

One solution I found on the web is as follows:

1) define a new command:

\newenvironment{ncases}[4]{%

\parbox{0.5\textwidth}{%

\begin{equation*}

#1=

\begin{cases}

\end{cases}

\end{equation*}}

\hfill

\parbox{1cm}{%

\begin{eqnarray}

{}\label{#3}\\

{}\label{#4}

\end{eqnarray}}}

Note: this command only supports two equations, if needed one can extend it easily to the number you want.

2) Use that command instead of the case command:

%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

\newenvironment{ncases}[4]{%

\parbox{0.5\textwidth}{%

\begin{equation*}

#1=

\begin{cases}

\end{cases}

\end{equation*}}

\hfill

\parbox{1cm}{%

\begin{eqnarray}

{}\label{#3}\\

{}\label{#4}

\end{eqnarray}}}

\begin{document}

\begin{spacing}{1.1}

\begin{ncases}

{|x|}{x, &\text{if $x\ge 0$}\\

-x,&\text{if $x\le 0$}}{pos}{neg}

\end{ncases}

For if $x\ge 0$, then by Equation~\eqref{pos}, we have $|x|=x\ge0$ and

if $x\le0$, then by Equation~\eqref{neg}, we have $|x|=-x\ge0$.

\end{spacing}

\end{document}

Posted via email from Troy's posterous

[Book] Springer Handbook of Speech Processing

2011-03-30T14:28:00.001+08:00

This book has a quiet complete explanation over various speech related techniques.

http://www.springerlink.com/content/978-3-540-49125-5#section=147599&page=1&locus=-1

Posted via email from Troy's posterous

[Linux] Prevent a background process from being stopped after closing SSH client

2011-03-28T15:24:00.001+08:00

Just find that the job could be run in the background on the server without being killed after log off the SSH terminal. Great!

Moreover, it will send you an email after the job is done!

The wonderful command is:

the 'batch' command.

$ batch > mycommand -x arg1 -y arg2 -z arg3 > ^D

This stuffs it in to the background, and then mails the results to you. It's a part of cron.

Reference from: http://stackoverflow.com/questions/285015/linux-prevent-a-background-process-from-being-stopped-after-closing-ssh-client

Posted via email from Troy's posterous

[CALL] Mispronunciation detection based on cross-language phonological comparisons

2011-03-25T23:06:00.001+08:00

10.1.1.149.1270 (1).pdf Download this file

This paper presents a method using speech recognition with linguistic constrains to detect the mispronunciations.

Compared with the standard ASR system, which consists of Acoustic Model, Lexicon and Language Model, the system used for mispronunciation detection only modifies the lexicon to include the possible phoneme confusions for recognition.

The phoneme confusions are gained from cross language phonological comparisons by human beings.

Thus the recognized results would possibly have more errors detected, which are interpreted as mispronunciations.

Actually, we can do alignment instead of recognition for mispronunciation detection. As in learning, the text is known to the speakers and speakers are asked to utter the given sentences.

In this paper, the measures they used are:

1) correctness: the percentage of all correctly detected phones;

2) accuracy: taking account of insertion

3) agreement of the system detection results with human judgments.

Posted via email from Troy's posterous

[Vis&Speech] TIMIT phoneme distribution

2011-03-24T20:50:00.001+08:00

I like figures because "a figure is worth 1000 words".

In the illustration below, the distribution of the 39 phoneme on the standard TIMIT corpus are reflected by the size of the circles.

Meanwhile, the positions of each phoneme actually represented some kind of similarity, although it is not verified by linguistic experts, from what I learned in my English classes, some confusing phones are indeed clustered together.

Posted via email from Troy's posterous

[HTK] HLDA transform estimation for High dimensional features

2011-03-08T22:57:00.001+08:00

For the TANDEM system, we could either train a set of HMM models on the projected posterior features, we could also train them directly on the posteriors and then using HTK to do a HLDA projection to reduce the dimension of the feature.

However, there is a limitation to the dimension of the feature vectors used: the dimension should be smaller than 100.

This is due to the HMath.c file, in which, there are several functions have defined some local array with the fixed length 100.

For my copy of the v3.4.1 HTK, those are:

1424: float col[100];
1427: int n,i,j,perm[100];
1477: double col[100];
1480: int n,i,j,perm[100];
1506: double col[100];
1509: int n,i,perm[100];
1535: double col[100];
1538: int n,i,perm[100];

What needs to do is just increase the size of those local arrays.

Posted via email from Troy's posterous

[Speech] Lecture Notes for ASR from Theory to Practice

2011-03-08T10:12:00.001+08:00

lecture-09.pdf Download this file

In the slides, it provides a good list of available tools for automatic speech recognition and thoughts for the future of ASR techniques.

Posted via email from Troy's posterous

[Linguistic] Symbols for American English Vowel Sounds

2011-03-07T14:25:00.001+08:00

From:

http://www.utexas.edu/courses/linguistics/resources/phonetics/vowelmap/index.html

The position mapping of vowels:

From:

http://www.utexas.edu/courses/linguistics/resources/phonetics/vowelmap/index.html

The Sun name for the phoneme is the one used in speech recognition.

A typical dialect of American English has about 15 distinctive vowel sounds. Here their symbols are linked to Sun-style .au samples lifted from the

ibiblio (Sunsite) archive (where they are listed without the .au extension).

The first symbol is the International Phonetic Association (IPA) symbol for the sound. (For the diphthongs, the American style of transcription is to use a -y where the standard IPA uses a 'j'.)
The second is the Sun name for the phoneme sample (which is in most cases the same as the symbol used by First Byte in Monologue for Windows and its DOS forebears).
The third symbol is the ipa-ascii symbol (an alphabet for use on Usenet groups and email).
The fourth column has the symbol that Rsynth displays in its verbose mode.
The fifth column contains the SAMPA symbol--as you can see, the differences among these alphabets are minor.
Each row concludes with a key word for the sound.

	S u n	IPAascii	Rsynth	Sampa	KeyWord
Front Vowels
h i g h l o w	IY	i	i	i	beet
	IH	I	I	I	bit
	EY	eI	eI	e	bait
	EH	E	e	E	bet
	AE	&	&	{	at

	S u n	IPAascii	Rsynth	Sampa	KeyWord
Back Vowels
h i g h . l o w	UY	u	u	u	boot
	UH	U	U	U	book
	OW	oU	oU	o	boat
	AO	O	O	O	cause
	AA	a/A	A	A	cot ¹

	S u n	IPAascii	Rsynth	Sampa	KeyWord
Central Vowels
	AX	@	@	@	about
	AH	V	V	V	but²

	S u n	IPAascii	Rsynth	Sampa	KeyWord
Diphthongs
	AY	aI	aI	aI	bite
	OY	OI	OI	OI	boy
	AW	AU	aU	aU	bough

Some links:
http://faculty.washington.edu/dillon/PhonResources/PhonResources.html#Recog
http://www.ibiblio.org/sounds/phonemes/

symbols for american english vowel sounds.pdf View this on Posterous

vowel mapping.pdf View this on Posterous

Posted via email from Troy's posterous

[Cocoa] A simple audio recording and playing demo using AudioQueue

2011-03-03T20:05:00.001+08:00

AudioRecord.zip View this on Posterous

This code is based on the Sample project SpeakHere of Apple developer site.

Posted via email from Troy's posterous

[Conference] Workshops related to Unsupervised learning and transfer learning

2011-03-01T20:13:00.001+08:00

Download now or preview on posterous

NIPS10_TLWS_Huang.pdf (114 KB)

ICML workshop on Transfer Learning:

http://clopinet.com/isabelle/Projects/ICML2011/

There is also a list of past papers for Transfer learning on that page.

Unsupervised and Transfer Learning Challenge

http://www.causality.inf.ethz.ch/unsupervised-learning.php?page=synopsis#cont

NIPS 2010 Workshop on Transfer Learning Via Rich Generative Models

Two interesting papers on this NIPS workshop:

http://www.mit.edu/~rsalakhu/workshop_nips2010/papers/NIPS10_TLWS_Huang.pdf

http://www.mit.edu/~rsalakhu/workshop_nips2010/papers/Piyush_mixture_nipsworkshop.pdf

Posted via email from Troy's posterous

[Misc] Mixing Objective-C and C++

2011-03-01T12:59:00.001+08:00

To use the C++ code in Objective-C, we can implement C++ class in pure C++. When using those C++ classes in Objective-C, any file having C++ types must have the extension of ".mm". Thus it seems not suitable to include C++ headers in Objective-C headers, although they are all ".h" files.

Posted via email from Troy's posterous

Interspeech proceeding index

2011-02-28T13:31:00.001+08:00

Download now or preview on posterous

INTERSPEECH proeeding index.pdf (497 KB)

Posted via email from Troy's posterous

[OpenGL] FreeType and FTGL for Xcode

2011-02-27T20:46:00.001+08:00

1) Download FreeType from: http://www.freetype.org and FTGL source from: http://sourceforge.net/projects/ftgl/

2) Extract the source and open a Terminal to navigate under the corresponding folder

3) For each do following steps in Terminal:

a) ./configure

b) make

4) find the libfreetype.a library file, for my case, it is under the folder of objs/.libs

5) copy the libfreetype.a and the include folder to a place, these are the files needed for apps;

6) Similarly for the FTGL, do ./configure, make and find the libftgl.a which is under src/.libs/

7) copy the libftgl.a and the folder FTGL under the src folder for future use.

For testing,

In Xcode, create a command-line app.

Copy the code from FTGL's demo folder's c-demo.c to the main.c of the newly created Xcode project.

drag libftgl.a, libfreetype.a and the header folders FTGL, freetype and the f2build.h to the project.

To compile the program we also need the "config.h" file from the FTGL's folder and add the OpenGL and GLUT frameworks.

Finally the result is :

Posted via email from Troy's posterous

[misc] speech recognition in real world - for fun

2011-02-24T15:47:00.001+08:00

Scotish Irish English American Eleven How you say eleven Sup Mee x Suck my willy :P Open The Door Pleasee :D Going Upp :P

Posted via email from Troy's posterous

[HMM] The GCONST in HTK's HMM definition file

2011-02-22T12:15:00.001+08:00

The GCONST value is a precomputed constant for the log likelihood of the Gaussian, i.e. the log of the denominator of the Gaussian formula:

The stored GCONST is:

where for the diagonal covariance matrix, the determinant is just the product of the diagonal elements.

Posted via email from Troy's posterous

[ML] Unsupervised Learning

2011-02-20T16:13:00.001+08:00

Download now or preview on posterous

fulltext (7).pdf (539 KB)

Unsupervised Learning lecture notes from MLSS.

Posted via email from Troy's posterous

[ML] Basics about Monte Carlo Methods

2011-02-20T11:23:00.001+08:00

Download now or preview on posterous

fulltext (6).pdf (1545 KB)

This is the lecture notes from MLSS for Monte Carlo Methods.

Posted via email from Troy's posterous

[MLSS] Singapore 2011

2011-02-20T11:16:00.001+08:00

http://mlss2011.comp.nus.edu.sg/index.php?n=Main.HomePage

Machine Learning Summer School

13-17 June 2011, Singapore

The home page for the MLSS world series is http://www.mlss.cc/

Lecture notes for previous events: http://www.springerlink.com/content/lrh41y849xdh/

Posted via email from Troy's posterous

[NN] A guide to recurrent neural network and backpropagation

2011-02-17T10:39:00.001+08:00

This doc is a rather simple introduction to Recurrent Neural Network giving a basic understanding of RNN.

Modeling time series, we can use either feedforward networks or RNN:

1) Adopting temporal windows on the input feature for feed forward NN, i.e. the tapped delay line memory in this paper

2) RNN is inherently capable of modeling the temporal information in the input sequence due to its recurrent characteristic

One major concern is the simple RNN only make use of the previous input information, which may not be sufficient especially for speech signals.

To incorporate more temporal information a complex system is required and the structure after unfolding is illustrated below:

Will this more complex system beat the feed forward NN with similar model complexity?

Another issue is the RNN only utilizes the historical information in the time series, while for windowing in feed forward NN, we can actually employ both the history and future information for current frame prediction.

However, feed forward NN process the current input and the context information with the same set of model parameter, which may not be the best way.

Maybe the first thing to try is just to experiment with RNN see whether it gives any promising results.

Download now or preview on posterous

10.1.1.3.9311.pdf (137 KB)

Posted via email from Troy's posterous

Introduction To Bayesian Inference

2011-02-16T15:36:00.001+08:00

See Also:

Launch in a standalone WM Player
Switch to Windows Media Player

Download slides: mlss09uk_bishop_ibi_01.pdf (2.8 MB)

View slides

Streaming Video Help

Windows Media Player Firefox Plugin - Download

via videolectures.net

In the progress of learning his wonderful book Pattern Recognition and Machine Learning

Posted via email from Troy's posterous

Probabilistic Models for Computational Linguistics

2011-02-16T15:35:00.001+08:00

See Also:

Launch in a standalone WM Player
Switch to Windows Media Player

Streaming Video Help

Windows Media Player Firefox Plugin - Download

via videolectures.net

Time to learn more basic and fundamental stuffs for my research

Posted via email from Troy's posterous

[Conference] PASCAL CHiME

2011-01-27T08:37:00.001+08:00

PASCAL CHiME Speech Separation and Recognition Challenge

Deadline: April 14, 2011
Workshop: September 1, 2011, Florence, Italy

http://www.dcs.shef.ac.uk/spandh/chime/challenge.html

Posted via email from Troy's posterous

[Speech] FFT algorithm

2011-01-26T16:13:00.001+08:00

The FFT algorithm implemented in the HTK is the Decimation-In-Time FFT algorithm as detailed explained in the attached file. The computation could be reflected in the following flow graph ( 8-poing DFT ):

A minor difference is that in the HTK implementation, the Wn=exp(j * 2 * pi / n), while in common the Wn adopted is Wn=exp(-j * 2 * pi).

The above FFT algorithm only deals with complex numbers. To do DFT on the real valued signal sequence we need a real version of the FFT, for which we could utilize the complex FFT. In the second document from Texas Instrument, they give a fast implementation of the FFT for real valued sequences ( on page 12-19). The algorithm is illustrated briefly in following figure:

Download now or preview on posterous

Decimation in time FFT algorithm.pdf (735 KB)

Download now or preview on posterous

DSP_FOR_FFT_COMPUTATION.PDF (389 KB)

Posted via email from Troy's posterous

[Learning] Unsupervised Learning

2011-01-21T11:10:00.001+08:00

Download now or preview on posterous

ul.pdf (362 KB)

Posted via email from Troy's posterous

[Speech] Front end analysis of speech recognition: A review

2011-01-20T17:43:00.001+08:00

Download now or preview on posterous

front end analysis of speech recognition_a review.pdf (2930 KB)

Posted via email from Troy's posterous

[Machine Learning] GTM: The Generative Topographic Mapping

2011-01-20T16:02:00.001+08:00

Download now or preview on posterous

GTM_The generative topographic mapping.pdf (463 KB)

Posted via email from Troy's posterous

[Misc] Command Line Keyboard Shortcuts for Mac OS X

2011-01-20T10:53:00.001+08:00

From: http://osxdaily.com/2006/12/19/command-line-keyboard-shortcuts-for-mac-os-x/

The command line in Mac OS X can be a very powerful and fun tool, so it’s good to know how to maneuver around if you find yourself in it. By default, the Mac OS X Terminal uses the Bash shell, which is what these keyboard shortcuts are intended for. So if you’re ready to get your feet wet, open up the Terminal and try these shortcuts out, they’re sure to make your command line life easier. The list isn’t too crazy so you should be able to try all these out within a minute or two, have fun:

Ctrl + A	Go to the beginning of the line you are currently typing on
Ctrl + E	Go to the end of the line you are currently typing on
Ctrl + L	Clears the Screen, similar to the clear command
Ctrl + U	Clears the line before the cursor position. If you are at the end of the line, clears the entire line.
Ctrl + H	Same as backspace
Ctrl + R	Let’s you search through previously used commands
Ctrl + C	Kill whatever you are running
Ctrl + D	Exit the current shell
Ctrl + Z	Puts whatever you are running into a suspended background process. fg restores it.
Ctrl + W	Delete the word before the cursor
Ctrl + K	Clear the line after the cursor
Ctrl + T	Swap the last two characters before the cursor
Esc + T	Swap the last two words before the cursor

Posted via email from Troy's posterous

[Tool] HMM Toolkit STK from Speech@FIT

2011-01-19T11:17:00.001+08:00

http://www.fit.vutbr.cz/research/groups/speech/sw/stk.html

Simple example scripts for MMI, MPE training.

Posted via email from Troy's posterous

[Speech] Hierarchical structures of neural networks for phoneme recognition

2011-01-19T10:52:00.001+08:00

Four Neural Network based phoneme recognition systems are investigated:

a) the TRAPs system (Fig. 1a) - separate networks for processing of speech in frequency bands;

b) the split temporal context (STC) system (Fig. 1b) - separate networks for processing of blocks of spectral vectors;

c) combination of both (Fig. 1c) - split in both frequency and time.

d) Tandem of two networks, the frond-end network is trained in classical ways and the back-end is trained on the combination of the front-end's posteriors and original features.

The assumptions for those systems are:

a) Independent processing of speech in critical bands;

b) Independent processing of parts of phonemes;

c) both a) and b).

Phoneme strings are basic representation for automatic language recognition and it is proved that language recognition results are highly correlated with phoneme recognition results. Phoneme posteriors are useful representation for acoustic keyword search, they contain enough information to distinguish among all words and they are small enough to store compared for example to the size of posteriors from context dependent Gaussian Mixture Models.

Two ways to provide additional information for NN training:

i) windowing, multiple frames context window, hamming window to emphasis the central frame;

ii) output representation: some improvements have been observed when a net was trained for multiple tasks in the same time.

A special Phoneme set mapping adopted in this paper is they merged closures with burst instead of with silence (bcl b -> b not bcl b -> pau b). It is believed that this mapping is more appropriate for features which use a longer temporal context.

The number of neurons in hidden layer of neural networks was increased until the saturation of phoneme error rate (PER) was observed. The obtained number of hidden layer neurons was approximately 500.

Table 1 shows the superiority of long Mel-bank energies but also great improvement coming from three state model. ( Block of 31 vectors of mel-bank energies (MBE) = 310 ms, Temporal trajectories in bands were weighted by Hamming window and down-sampled by DCT to 11 coefficients. )

The final best PER reported in this paper is using the 5-block STC system with bigram LM as shown in following table:

Download now or preview on posterous

ICASSP2006_Schwarz_PhnRec.pdf (75 KB)

Posted via email from Troy's posterous

[Machine Learning] Machine Learning Summer School 2008 - Kioloa

2011-01-17T20:37:00.001+08:00

http://videolectures.net/mlss08au_kioloa/

Posted via email from Troy's posterous

Monte Carlo Simulation for Statistical Inference, Model Selection and Decision Making

2011-01-17T20:27:00.000+08:00

Monte Carlo Simulation for Statistical Inference, Model Selection and Decision Making

2011-01-17T20:25:00.001+08:00

See Also:

Launch in a standalone WM Player
Switch to Windows Media Player

Download slides: mlss08au_freitas_asm.pdf (14.4 MB)

Streaming Video Help

Windows Media Player Firefox Plugin - Download

via videolectures.net

Posted via email from Troy's posterous

[Tool] TexPoint - A latex plugin for Microsoft Office

2011-01-17T17:20:00.001+08:00

http://texpoint.necula.org/index.html

Posted via email from Troy's posterous

[Speech] Mixture Density Network (Technical Report)

2011-01-17T16:52:00.001+08:00

Download now or preview on posterous

Mixture Density Network.pdf (428 KB)

Posted via email from Troy's posterous

[Speech] A trajectory density mixture network for acoustic articulatory inversion mapping

2011-01-17T16:49:00.001+08:00

Download now or preview on posterous

A trajectory mixture density network for acoustic articulatory inversion mapping.pdf (113 KB)

Mixture Density Network

PRML Chapter 5.

http://www.cedar.buffalo.edu/~srihari/CSE574/Chap5/Chap5.6-MixDensityNetworks.pdf

Posted via email from Troy's posterous

Extracting and Composing Robust Features with Denoising Autoencoders

2011-01-13T21:36:00.000+08:00

Extracting and Composing Robust Features with Denoising Autoencoders

[DBN] Learning multiple layers of Features from tiny images

2011-01-07T15:32:00.001+08:00

Download now or preview on posterous

learning-features-2009-TR.pdf (4107 KB)

In this paper, there are more detailed equation derivations for RBM, especially the Gaussian-Bernoulli RBM.

Posted via email from Troy's posterous

Matlab v7.3 mat file and python

2011-01-05T20:27:00.001+08:00

From: http://mloss.org/community/blog/2009/nov/19/matlabtm-73-file-format-is-actually-hdf5-and-can-b/

It looks like that matlab version 7.3 and later are capable of writing out objects in the so called matlab 7.3 file format. While at first glance it looks like another proprietary format - it seems to be in fact the Hierarchical Data Format version 5 or in short hdf5.

So you can do all sorts of neat things:

Lets create some matrix in matlab first and save it:

>> x=[[1,2,3];[4,5,6];[7,8,9]] x = 1 2 3 4 5 6 7 8 9 >> save -v7.3 x.mat x

Lets investigate that file from the shell:

$ h5ls x.mat x Dataset {3, 3} $ h5dump x.mat HDF5 "x.mat" { GROUP "/" { DATASET "x" { DATATYPE H5T_IEEE_F64LE DATASPACE SIMPLE { ( 3, 3 ) / ( 3, 3 ) } DATA { (0,0): 1, 4, 7, (1,0): 2, 5, 8, (2,0): 3, 6, 9 } ATTRIBUTE "MATLAB_class" { DATATYPE H5T_STRING { STRSIZE 6; STRPAD H5T_STR_NULLTERM; CSET H5T_CSET_ASCII; CTYPE H5T_C_S1; } DATASPACE SCALAR DATA { (0): "double" } } } } }

And load it from python:

>>> import h5py >>> import numpy >>> f = h5py.File('x.mat') >>> x=f["x"] >>> x <HDF5 dataset "x": shape (3, 3), type "<f8"> >>> numpy.array(x) array([[ 1., 4., 7.], [ 2., 5., 8.], [ 3., 6., 9.]])

So it seems actually to be a good idea to use matlab's 7.3 format for interoperability.

Posted via email from Troy's posterous

[Deep] Deep learning tutorial on MLSS 2010

2010-12-28T16:34:00.001+08:00

Download now or preview on posterous

Frean1.pdf (11467 KB)

More about MLSS 2010 could be found http://mlss10.rsise.anu.edu.au/proceedings

Posted via email from Troy's posterous

[ASR] Still Tandem systems

2010-12-23T12:05:00.001+08:00

Download now or preview on posterous

combining phonetic attributes using conditional random field.pdf (112 KB)

The attached paper models phonetic attributes with CRF models. However, the sentence interests me most is the following one:

As described in [1] we used the linear output of the MLPs with a KL transform applied to them to decorrelate the features, as this gave the best results for the HMM system.

[1] H. Hermansky, D. Ellis, and S. Sharma, "Tandem connectionist feature stream extraction for conventional HMM systems", in Proc. of the ICASSP 2000.

Maybe sometime we could also try to refine the posterior features from NN for HMM systems.

Posted via email from Troy's posterous

[Linux] Mac configure type

2010-12-20T20:57:00.001+08:00

From: http://macosx.com/forums/unix-x11/29486-guessing-host-type-since-10-2-4-a.html

http://hints.macworld.com/article.php?story=2001041322081284

In many cases the build step involves running a configure script. Occasionally you'll run a configure script that gives error message stating that the host type cannot be determined. In such a case, you can usually either specify the configure option --host=powerpc-apple-darwin6.4 as I've done in the examples below. (Note in this case, we are using darwin6.4. To determine the correct release of darwin, enter the command uname -r.) You can alternatively copy two files into the build directory, that is, into the same directory that contains the script configure.

Code:

cp /usr/share/libtool/config.guess . cp /usr/share/libtool/config.sub .

Occasionally, the configure option --build=powerpc-apple-darwin6.4 is needed to produce the desired result (again assuming we're using darwin release 6.4).

In some cases, there are additional steps to be performed. ...

Posted via email from Troy's posterous

[Info] Machine Learning Summer School 2009

2010-12-19T23:19:00.001+08:00

http://mlg.eng.cam.ac.uk/mlss09/

Posted via email from Troy's posterous

[Basic] Whitening

2010-12-16T17:17:00.001+08:00

Download now or preview on posterous

whiten.pdf (37 KB)

From: http://cis.legacy.ics.tkk.fi/aapo/papers/IJCNN99_tutorialweb/node26.html

Another useful preprocessing strategy in ICA is to first whiten the observed variables. This means that before the application of the ICA algorithm (and after centering), we transform the observed vector linearly so that we obtain a new vector which is white, i.e. its components are uncorrelated and their variances equal unity. In other words, the covariance matrix of equals the identity matrix:

(30)

The whitening transformation is always possible. One popular method for whitening is to use the eigen-value decomposition (EVD) of the covariance matrix, where is the orthogonal matrix of eigenvectors of and is the diagonal matrix of its eigenvalues, . Note that can be estimated in a standard way from the available sample . Whitening can now be done by

(31)

where the matrix is computed by a simple component-wise operation as . It is easy to check that now .

Whitening transforms the mixing matrix into a new one, . We have from (4) and (34):

(32)

The utility of whitening resides in the fact that the new mixing matrix is orthogonal. This can be seen from

(33)

Here we see that whitening reduces the number of parameters to be estimated. Instead of having to estimate the n² parameters that are the elements of the original matrix , we only need to estimate the new, orthogonal mixing matrix . An orthogonal matrix contains n(n-1)/2degrees of freedom. For example, in two dimensions, an orthogonal transformation is determined by a single angle parameter. In larger dimensions, an orthogonal matrix contains only about half of the number of parameters of an arbitrary matrix. Thus one can say that whitening solves half of the problem of ICA. Because whitening is a very simple and standard procedure, much simpler than any ICA algorithms, it is a good idea to reduce the complexity of the problem this way.

It may also be quite useful to reduce the dimension of the data at the same time as we do the whitening. Then we look at the eigenvalues d_j of and discard those that are too small, as is often done in the statistical technique of principal component analysis. This has often the effect of reducing noise. Moreover, dimension reduction prevents overlearning, which can sometimes be observed in ICA [26].

A graphical illustration of the effect of whitening can be seen in Figure 10, in which the data in Figure 6 has been whitened. The square defining the distribution is now clearly a rotated version of the original square in Figure 5. All that is left is the estimation of a single angle that gives the rotation.

Figure 10: The joint distribution of the whitened mixtures.

In the rest of this tutorial, we assume that the data has been preprocessed by centering and whitening. For simplicity of notation, we denote the preprocessed data just by , and the transformed mixing matrix by , omitting the tildes.

Posted via email from Troy's posterous

[Basic] Cross-Entropy Criterion

2010-12-13T11:39:00.001+08:00

Download now or preview on posterous

Kullback-Leibler divergence.pdf (281 KB)

Cross-Entropy Criterion is actually the Kullback Leibler Divergence.

KL divergence is a non-symmetric measure of the difference between two probability distributions P and Q. KL measures the expected number of extra bits required to code samples from P when using a code based on Q, rather than using a code based on P. Thus, P is the true distribution and Q is the estimated distribution. Typically P represent the "true" distribution of data, observations, or a precise calculated theoretical distribution. The measure Q typically represent a theory, model, description, or approximation of P.

Although it is often intuited as a distance metric, the KL divergence is not a true metric - for example, it's not symmetric: the KL from P to Q is not necessarily the same as the KL from Q to P.

For probability distributions P and Q of a discrete random variable their KL divergence is defined to be:

D_KL(P||Q)=\sum{ P(i) log[ P(i)/Q(i) ] }

In words, it is the average of the logarithmic difference between the probabilities P and Q, where the average is taken using the probabilities P. The KL divergence is only defined if P and Q both sum to 1 and if Q(i) > 0 for any i such that P(i) > 0. If the quantity 0 log 0 appears in the formula, it is interpreted as zero.

The attached document is the explanation for KL divergence from wikipedia.

Posted via email from Troy's posterous

Printing in NUS SoC

2010-12-10T14:28:00.001+08:00

From: http://opensource.nus.edu.sg/wiki/index.php/Printing_in_NUS_SoC

[hide]

1 Printing in the School of Computing

Printing in the School of Computing

Printing in SoC has long been a nightmare for users of "unsupported systems" like us. Fortunately, the most difficult part is actually *figuring out* how to get printing working on Linux, and since we already have, the steps you need to take to enable printing on your Linuxbox is relatively simple, especially if you're using a fairly recent and popular distro.

Here are detailed steps for recent versions of Ubuntu and OpenSUSE. They should be fairly similar in comparable distros.

Printing in Ubuntu

Since around the time of Hardy, everything you need to print to the SoC printers is available out-the-box on an Ubuntu system. To start printing, you simply need to configure a printer as follows:

Step 1

Make sure you are either connected to PEAP or to SoCVPN.

Go to System > Administration > Printing, key in your password, and hit the "New" button on the window that appears.

Step 2

Select Network Printer > Windows Printer via Samba on the left-hand column.

On the right, enter the following in the textfield under "SMB Printer"

nusstu/nts27.comp.nus.edu.sg/psts-dx

You can replace 'psts-dx' above with the name of whatever target printer you desire to add.

Check "Set authentication details now" and key in your NUSNET credentials.

You can hit "verify" to see if your connection is successful. Hit "Forward" when you're done.

Step 3

Select the printer manufacturer here. For the COM1 printers, pick "Lexmark". Then hit "Forward".

Step 4

Select the printer model here. For the COM1 printers, pick "T642".

Step 5

Optionally Key in some meta information for your new printer.

Hit "Apply" and you're done!!

Print a test page to boost your ego.

Printing in Kubuntu

IdyllicTux has a Video Tutorial on printing on Kubuntu. Steps are highly similar to those for Ubuntu.

Printing in OpenSUSE

This is based on the KDE version. Steps for OpenSUSE Gnome should be similar.

Step 1

Make sure you are either connected to PEAP or to SoCVPN.

Bring up the Yast Administrator Settings Panel (Application > System > Administrator Settings) and select "Printer"

In the window that opens, make sure "Printer configurations" is selected on the left-hand panel, then click "Add" near the bottom-center.

Step 2

Click "Connection Wizard" at the top-right.

In the page that results, Select "Print via Server Machine > Microsoft Windows/SAMBA (SMB/CIFS)" on the left-hand panel.

On the right, key in

"nts27.comp.nus.edu.sg" for Server (NetBIOS Host Name)
Your target printer name for Printer (Share Name), eg. psts-dx
"nusstu" for Workgroup (Domain Name)
Your NUSNET credentials under "Authenticate As"
The printer manufacturer - "Lexmark" for the COM1 printers on the last drop-down.

Test the connection if you like. Hit "ok" when you're done.

Step 3

Under "Assign Driver", look for and select the driver for the target printer. For the COM1 printers, select the entry beginning with "Lexmark T642".

If you cannot find this in the list, you may need to install a driver package. Click "Add Driver" to do so. The following page should appear:

Check off the driver packages you would like to install. "cups-drivers" should contain the T642 driver for the COM1 printers, so check it if you haven't. Hit "ok" to install the driver packages.

You will be returned to the original page once you're done, and you should now find your driver in the list. Select it and hit "ok".

You're done!! Print a test page to boost your ego.

A word on changing password

The Yast printer configuration tool is a little unwieldy, and shockingly does not appear to have a direct way of changing your authentication password. When you change your NUSNET password, simply delete your existing printing configurations and re-add them following the steps above.

Posted via email from Troy's posterous

[Ogre] SDK Install

2010-12-06T18:52:00.001+08:00

After dragging the SDK folder to local disk, before open the xcode project, be sure to run following command to make the xcode project adjusted to the current environment.

cmake -GXcode .

Posted via email from Troy's posterous

3D Little Human Figures Clipart Big Collection

2010-12-04T20:05:00.001+08:00

3D Little Human Figures Clipart – Big Collection
115 jpg | Up to 9001*6000 pix | 300 dpi | 282 Mb rar

Download:

Part 1 http://depositfiles.com/en/files/a9wxaatgz
Part 2 http://depositfiles.com/en/files/7f2ftolkj
Part 3 http://depositfiles.com/en/files/bozjkzeza

From: http://lordofdesign.com/3d-little-human-figures-clipart-big-collection/

Posted via email from Troy's posterous

[Speech] Power spectrum

2010-12-01T18:30:00.001+08:00

Download now or preview on posterous

Power.pdf (45 KB)

Posted via email from Troy's posterous

[Latex] LaTex: Fixing Wrong Figure Numbers

2010-12-01T13:52:00.001+08:00

From: http://www.terminally-incoherent.com/blog/2007/04/14/latex-fixing-wrong-figure-numbers/

ONE SENTENCE: PUT THE LABEL COMMAND AFTER CAPTION COMMAND FOR FIGURES.

What I tell you right now may save you hours of extensive debugging, cursing under your breath, commenting out custom code dealing with figure layout and much frustration. Whenever you use figures, always (and I mean ALWAYS EVER FOREVER ALWAYS) put `\caption` first, and `\label`second like this:

\begin{figure}[htp]         \centering                 \includegraphics{image.eps}         \caption{Some Image}         \label{fig:some-image} \end{figure}

If you put the \label above \caption you will run into trouble when referencing figures inside subsections. In my case, the caption underneath the figure would say Fig. 4.2 but the output of \ref would be 4.3.10 because somehow it was picking up the section numbers wrong. The whole damn chapter 4 had the caption/label pairs flipped – but the rest of the document was fine. I have no clue what possessed me to write it this way.

Posted via email from Troy's posterous

[Feature] SIFT

2010-11-30T23:02:00.001+08:00

Download now or preview on posterous

tutSIFT04.pdf (3040 KB)

In spite of significant progress in automatic speech recognition over the years, robustness still appears to be a stumbling block. Current commercial products are quite sensitive to changes in recording device, to acoustic clutter in the form of additional speech signals, and so on. The goal of replicating human performance in a machine remains far from sight.

Scale Invariant Feature Transform (SIFT) is an approach for detecting and extracting local feature descriptors that are reasonably invariant to changes in illumination, image noise, rotation, scaling, and small changes in viewpoint.

Detection stages for SIFT features:

1) Scale-space extrema detection

Interest points for SIFT features correspond to local extrema of difference-of-Gaussian filters at different scales.

Interest points (called keypoints in the SIFT framework) are identified as local maxima or minma of the DoG (difference of Gaussian) images across scales. Each pixel in the DoG images is compared to its 8 neighbors at the same scale, plus the 9 corresponding neighbors at neighboring scales. If the pixel is a local maximum or minimum, it is selected as a candidate keypoint.

For each candidate keypoint:

- Interpolation of nearby data is used to accurately determine its position;

- Keypoints with low contrast are removed;

- Responses along edges are eliminated;

- The keypoint is assigned an orientation.

To determine the keypoint orientation, a gradient orientation histogram is computed in the neighborhood of the keypoint (using the Gaussian image at the closest scale to the keypoint's scale). The contribution of each neighboring pixel is weighted by the gradient magnitude and a Gaussian window with a theta that is 1.5 times the scale of the keypoint.

Peaks in the histogram correspond to dominant orientations. A separate keypoint is created for the direction corresponding to the histogram maximum, and any other direction within 80% of the maximum value.

All the properties of the keypoint are measured relative to the keypoint orientation, this provides invariance to rotation.

2) Key point localization

3) Orientation assignment

4) Generation of keypoint descriptors

Posted via email from Troy's posterous

[Feature] Speech Recognition with localized time-frequency pattern detectors

2010-11-30T12:11:00.001+08:00

Characteristics of the localized time-frequency features:

1) Local in frequency domain, not like MFCC, each feature is affected by all the frequencies;

2) Temporal dynamics, modeling long and variable time durations, while in MFCC, the features are all short time and fixed duration.

In this paper, the set of filters adopted is very simple, and are essentially basic edge detectors taking only values +1 and -1. The selection includes vertical edges (of varying frequency span and temporal duration) for onsets and offsets; wide horizontal edges for frication cutoffs; and horizontal edges tilted at various slopes to model formant transitions. The choices for the ranges of the various parameters were made based on acoustic phonetic knowledge, such as typical formant bandwidths, average phone durations, typical rates of formant movement, etc.[Book: Acoustic Phonetics]

With these filters, the features are computed as follows:

For each filter,

1) centering it at a particular frequency, and convolving with the whole spectrogram along the time axis for that specific frequency;

2) for each frequency value, we could get a time series of the convolution sums;

3) to reduce the dimension of features, the convolution sums are down sampled over a 16*32 point grid.

4) The 32 frequency points are taken linearly between 0 and 4kHz.

5) The 16 time points (which is specific to their task) is the centers of the 16 states (for the HMM word model) in the state alignment.

Thus, a feature refers to both the filter shape, and the time-frequency point (16*32 point grid).

In this paper, the task is to classify isolated digits. Each spectrogram is one single digit. The feature computed is per digit, thus is global to the target class.

###################

The problem with current frame based features is its non-localization, make it difficult to modeling speaker variabilities.

As shown in following figure:

Download now or preview on posterous

Schutte, Glass_2008_Speech recognition with localized time-frequency pattern detectors.pdf (374 KB)

Posted via email from Troy's posterous

[Speech] Phonetic cues

2010-11-29T20:28:00.001+08:00

Download now or preview on posterous

Glass, Schutte_2009_Parts-based models and local features for automatic speech recognition.pdf (2115 KB)

While automatic speech recognition systems have steadily improved and are now in widespread use, their accuracy continues to lag behind human performance, particularly in adverse conditions.

There has been much progress and ASR technology is now in widespread use; however, there is still a considerable gap between human and machine performance, particularly in adverse conditions.

How human evolution makes humans different from machines in perceiving speech signals?

What's the major differences between humans and machine when processing speech signals? And which are the crucial ones?

The parts-based model (PBM), based on previous work in machine vision, uses graphical models to represent speech with a deformable template of spectro-temporally localized "parts", as opposed to modeling speech as a sequence of fixed spectral profiles.

Perhaps most importantly, ASR systems have benefited greatly from general improvements in computer technology. The availability of very large datasets and the ability to utilize them for training models has been very beneficial. Also, with ever increasing computing power, more powerful search techniques can be utilized during recognition.

Reaching the ultimate goal of ASR - human-level (or beyond) performance in all conditions and on all tasks - will require investigating other regions of this landscape, even if doing so results in back-tracking in progress in the short-term.

Utilize the knowledge of acoustic phonetics and human speech perception in speech recognition.

We will argue that well known phonetic cues crucial to human speech perception are not modeled effectively in standard ASR systems, and that there is a benefit to model such cues explicitly, rather than implicitly.

The "glimpsing" model of speech perception suggests that humans can robustly decode noise-corrupted speech by taking advantage of local T-F regions having high SNR, and is supported by empirical evidence and computational models.

Auditory Neuroscience, Tonotopic maps. Also, recent research seeking to characterize the behavior of individual neurons in the mammalian auditory cortex has resulted in models in which cortical neurons act as localized spectro-temporal pattern detectors. ( represented by their so-called spectro-temporal receptive filed, or STRF).

+ Localized T-F pattern detectors

+ Explicit modeling of phonetic cues

Corpora:

ISOLET http://www.icsi.berkeley.edu/speech/papers/eurospeech05-onset/isolet/

VCV http://www.odettes.dds.nl/challenge_IS08/downloads.html

acoustic ----------> phonetic cues ------------------> phonemes

Posted via email from Troy's posterous

[DBN] Learning rate for RBM training

2010-11-29T13:46:00.001+08:00

Download now or preview on posterous

learnrate.pdf (81 KB)

Posted via email from Troy's posterous

[Speech] Spectrogram

2010-11-25T15:21:00.001+08:00

From: http://www-3.unipv.it/cibra/edu_spectrogram_uk.html

To analyze sounds it is required to have an acoustic receiver (a microphone, an hydrophone or a vibration transducer) and an analyzer suitable for the frequencies of the signals we want to measure. Eventually, a recorder may allow to permanently store the sounds to allow later analyses or playbacks.

A spectrograph transforms sounds into images to make "visible", and thus measurable and comparable, sound features the human hear can't perceive. Spectrograms (also called sonograms or sonagrams) may show infrasounds, like those emitted by some large whales or by elephants, as well as ultrasounds, like those emitted by echolocating dolphins and by echolocating bats, but also emitted by insects and small rodents.

Spectrograms may reveal features, like fast frequency or amplitude modulations we can't hear even if they lie within our hearing frequency limits (30 Hz - 16 kHz). Spectrograms are widely used to show the features of animal voices, of the human voice and also of machinery noise.

A real-time spectrograph displays continuously the results of the analyses on the incoming sounds with a very small - often not perceivable - delay. This kind of instrumentation is very useful in field research because it allows to continuously monitor the sounds received by the sensors, to immediately evaluate their features, and to classify the received signals. A spectrograph can be dedicated instrument or a normal computer equipped with suitable hardware for receiving and digitizing sounds and a software to analyze sounds and convert them into a graphical representation.

Normally, a spectrogram represents the time on the x axis, frequency on the y axis and the amplitude of the signals by using a scale of grays or a scale of colours. In some applications, in particular those related with military uses, the x and y axes are swapped.

The quality and features of a spectrogram are controlled by a set of parameters. A default set can be used for generic display, but some parameters can be changed to optimize the display of specific features of the signals.
Also, by modifying the colour scale it is possible to optimize the display of the amplitude range of interest.

Posted via email from Troy's posterous

[News] ACMTech Nov.23

2010-11-23T10:04:00.001+08:00

AT&T Ups the Ante in Speech Recognition
CNet (11/18/10) Marguerite Reardon

AT&T says it has devised technologies to boost the accuracy of speech and language recognition technology as well as broaden voice activation to other modes of communication. AT&T's Watson technology platform is a cloud-based system of services that identifies words as well as interprets meaning and contexts to make results more accurate. AT&T recently demonstrated various technologies such as the iRemote, an application that transforms smartphones into voice-activated TV remotes that let users speak natural sentences asking to search for specific programs, actors, or genres. Most voice-activated remotes respond to prerecorded commands, but the iRemote not only recognizes words, but also employs other language precepts such as syntax and semantics to interpret and comprehend the request's meaning. AT&T also is working on voice technology that mimics natural voices through its AT&T Natural Voices technology, which builds on text-to-speech technology to enable any message to be spoken in various languages, including English, French, Italian, German, or Spanish when text is processed via the AT&T cloud-based service. The technology accesses a database of recorded sounds that, when combined by algorithms, generate spoken phrases.
http://news.cnet.com/8301-30686_3-20023189-266.html

What If We Used Poetry to Teach Computers to Speak Better?
McGill University (11/17/10)

McGill University linguistics researcher Michael Wagner is studying how English and French speakers use acoustic cues to stress new information over old information. Finding evidence of a systematic difference in how the two languages use these cues could aid computer programmers in their effort to produce more realistic-sounding speech. Wagner is working with Harvard University's Katherine McCurdy to gain a better understanding of how people decide where to put emphasis. They recently published research that examined the use of identical rhymes in poetry in each language. The study found that even when repeated words differ in meaning and sound the same, the repeated information should be acoustically reduced as otherwise it will sound odd. "Voice synthesis has become quite impressive in terms of the pronunciation of individual words," Wagner says. "But when a computer 'speaks,' whole sentences still sound artificial because of the complicated way we put emphasis on parts of them, depending on context and what we want to get across." Wagner is now working on a model that better predicts where emphasis should fall in a sentence given the context of discourse.
http://www.eurekalert.org/pub_releases/2010-11/mu-wiw111710.php

Posted via email from Troy's posterous

Dream & Passion

WSJ Setup

[HTK] Chinese Encoding

[HTK] Chinese encoding

[HTK] Increase HTK feature dimension limit

Solving iMessage or FaceTime waiting for activation problem

Developing a Static Library and Incorporating It in Your Application in Xcode4

Developing a Static Library and Incorporating It in Your Application

Tutorial: Code Sharing Via Static Libraries And Cross-Project References

Tutorial: Code Sharing Via Static Libraries And Cross-Project References

Wanted: An Elegant Way To Share Code Across Projects

Solution: Static Libraries Built On-Demand Via Xcode Cross-Project References

Implementing Cross-Project References

HDecode in lattice rescoring mode

[HTK] CreateHMM: multiple use of logical HMM name sp

PCA whitening

PCA again

Unity3D Resources

Resolving nvcc & gcc conflict ion for Theano

Papers on learning sparse Gaussian Precision Matrix

LangBrain Website

Graphical Gaussian Models for Genome Data

[Math] The Matrix Cookbook

HTK lattice format

A hierarchical context dependent neural network architecture for improved phone recognition

Dirichlet mixture models of neural net posteriors for hmm based speech recognition

Automatic speech recognition using hidden conditional neural fields

Slides for Introduction to Restricted Boltzmann Machine

[HTK] Compile HTK on 64bit machine

Ubuntu 9.04 nVidia Driver Screen Resolution Problem

[Matlab] staring error on Ubuntu - libc.so.6 not found

MATLAB with Ubuntu 11.04

[CUDA] Solving the problem of not supporting gcc-4.5 and above

Supervised Clustering

See Also:

[ASR] Langue model

[Tools] HVite for alignment

[Tools] Compiling Matlab .m Files

[ASR] A complete Matlab script for TIMIT phoneme recognition

[ASR] TIMIT information for phoneme recognition

[NN] NN output activations with cross entropy error function

[Business] Business Plan Formating

BUSINESS PLANFORMATTING

COMPANY PURPOSE

PROBLEM

SOLUTION

WHY NOW

MARKET SIZE

COMPETITION

PRODUCT

BUSINESS MODEL

TEAM

FINANCIALS

[Business] Business model

[ML] Some learnings

[DBN] Deep Machine Learning - A New Frontier in Artificial Intelligence Research

[DBN] Deep Learning and Its Applications to Signal and Information Processing

Pronunciation teaching - Internet Resources

[Latex] Simultaneous Equations in Latex

[Book] Springer Handbook of Speech Processing

[Linux] Prevent a background process from being stopped after closing SSH client

[CALL] Mispronunciation detection based on cross-language phonological comparisons

[Vis&Speech] TIMIT phoneme distribution

[HTK] HLDA transform estimation for High dimensional features

[Speech] Lecture Notes for ASR from Theory to Practice

[Linguistic] Symbols for American English Vowel Sounds

[Cocoa] A simple audio recording and playing demo using AudioQueue

[Conference] Workshops related to Unsupervised learning and transfer learning

Unsupervised and Transfer Learning Challenge

NIPS 2010 Workshop on Transfer Learning Via Rich Generative Models

[Misc] Mixing Objective-C and C++

Interspeech proceeding index

[OpenGL] FreeType and FTGL for Xcode

[misc] speech recognition in real world - for fun

[HMM] The GCONST in HTK's HMM definition file

[ML] Unsupervised Learning

[ML] Basics about Monte Carlo Methods

[MLSS] Singapore 2011

[NN] A guide to recurrent neural network and backpropagation

Introduction To Bayesian Inference

BUSINESS PLAN
FORMATTING

What I tell you right now may save you hours of extensive debugging, cursing under your breath, commenting out custom code dealing with figure layout and much frustration. Whenever you use figures, always (and I mean ALWAYS EVER FOREVER ALWAYS) put `\caption` first, and `\label`second like this: