Exploring deep neural networks and deep autoencoders in reverberant speech recognition
Masato Mimura, Shinsuke Sakai, Tatsuya Kawahara · 2014
We propose an approach to reverberant speech recognition adopting deep learning in front end as well as back end of the system. At the front end, we adopt a deep autoencoder (DAE) for enhancing the speech feature parameters, and speech recognition is performed using a DNN-HMM acoustic models at the back end. The system was evaluated on simulated and real reverberant speech data sets. On average, the DNN-HMM system trained on the multi-condition training data outperformed the MLLR-adapted GMM-HMM system trained on the same data. The feature enhancement with the DAE contributed to the improvement of recognition accuracy especially in more adverse conditions. We also performed an unsupervised adaptation of the DNN-HMM models to the test data enhanced by the DAE and achieved improvements in word accuracies in all reverberation conditions of the test data.