Three ways to adapt a CTS recognizer to unseen reverberated speech in BUT system for the ASpIRE challenge

speech recognition, reverberation, dereverberation, neural networks, DNN

We have presented our work towards the ASR of wide-band noisy reverberant speech in ASpIRE challenge. To solve this task, we have started with augmenting Fisher data with artificially noised and reverberated versions.

This paper describes several strategies tested in BUT’s submission to the IARPA ASpIRE challenge. The ASpIRE task was to develop an automatic speech recognition (ASR) system for wide-band noisy reverberant speech, while only clean CTS (Fisher) data was allowed for ASR training. To solve this task, we have started with augmenting Fisher data with artificially noised and reverberated versions. The most obvious adaptation was (1) to re-train the whole GMM/HMM-based ASR system. Then, two techniques were designed and tested to make the adaptation easier and overcome retraining the whole ASR on large amount of speech: (2) we trained a speech enhancement DNN (also called de-noising auto-encoder), and (3) we adapted the feature extraction based on stacked bottle-neck networks (SBN). While re-training the whole system works the best, only slightly inferior results were obtained with the autoencoder denoising followed by retraining of the first layers of the SBN hierarchy, letting most of the ASR system trained on clean Fisher unchanged. This shows a promising, efficient and fast way to port ASR systems to new conditions.