Enhanced Denoising Auto-Encoder for Robust Speech Recognition in Unseen Noise Conditions
Sonal Joshi, Ashish Kumar Panda, Biswajit Kumar Das · 2018
We present a robust front-end processing method for speech recognition in unseen noise conditions. Towards this end, we have investigated the efficacy of a Time Delay Neural Network based Denoising Auto-Encoder (TDNN-DAE) in seen and unseen noise conditions. We show that while the TDNN-DAE succeeds in improving the performance of the speech recognition by a large margin in seen noise conditions (noise encountered during decoding was used in the training of the TDNN-DAE), it fails to improve the performance in unseen noise conditions (noise encountered during decoding was not used in the training of the TDNN-DAE). To address this, we propose to pre-process the training input to the TDNN-DAE using an enhancement technique. In essence, the TDNN-DAE is being trained to address the residual noise left behind by the enhancement technique. For this task, we compare the performance of two enhancement techniques, namely Vector Taylor Series with Acoustic Masking (VTS-AM) and Spectral Subtraction (SS). We show that both these enhancement techniques improve the efficacy of the TDNN-DAE significantly in unseen noise conditions and that the VTS-AM enhanced TDNN-DAE outperforms the SS enhanced TDNN-DAE.