Application of multi-layer perceptron in estimating speech/noise characteristics for speech recognition in noisy environment

In this paper, we will consider the problem of speech recognition under noisy conditions. Multi-layer perceptron (MLP) based estimators estimating filter bank channel outputs are tested on a speaker independent isolated digit recognition task in white noise conditions. Local statistical information of the speech and the noise are estimated online and used as inputs to the estimators. The results are comparable to those obtained in clean condition at a moderate signal-to-noise ratio (SNR) of 20 dB. Substantial improvement is also obtained at lower SNRs. By carefully studying the results, it is noted that the MLP based estimators appear to perform poorly when there is virtually no detectable significant speech activity. As a result, a modified gain function is introduced. This improves the performance even further.