[Paper] Increased Robustness of Noisy Speech Features Using Neural Networks

This is an early paper (1999) from Nuance working on noise robust ASR.

The assumption in this paper is that the noises are additive and can be modeled simply by a bias vector either per frame or per utterance. 

This bias vector is estimated from neural networks using different knowledge sources. They found that estimating the bias vectors for log filterbanks is easier than cepstral vectors.

The paper could be found at

Increased Robustness of Noisy Speech Features Using Neural Networks