Integrating Denoising Autoencoder and Vector Taylor Series with Auditory Masking for Speech Recognition in Noisy Conditions

摘要

We propose a new front-end feature compensation technique to improve the performance of Automatic Speech Recognition (ASR) systems in noisy environments. First, a Time Delay Neural Network (TDNN) based Denoising Autoencoder (DAE) is considered to compensate the noisy features. The DAE provides good gain in performance when it has been trained using the noise present in the test utterances (“seen” conditions). However, if the noise present in the test utterance is different to what was used in the training of the DAE (“un-seen” conditions), then the performance degrades to a great extent. To improve the ASR performance in such unseen conditions, a model compensation technique, namely the Vector Taylor Series with Auditory Masking (VTS-AM) is used. We propose a new Signal-to-Noise Ratio (SNR) based measure, which can reliably choose the type of compensation to be used for best performance gain. We show that the proposed technique improves the ASR performance significantly on noise corrupted TIMIT and Librispeech databases.