Comparing Front-End Enhancement Techniques and Multiconditioned Training for Robust Automatic Speech Recognition

We present comparison of various front-end enhancement techniques and multiconditioned training for robust Automatic Speech Recognition (ASR) for additive noise. We compare De-noising Autoencoders (DAEs) based on Deep Neural Network (DNN), Time-Delay Neural Network (TDNN) architecture, and Time-Frequency (T-F) masking based DNN based front-ends. We train these front-ends and evaluate their performance on various seen/unseen noise conditions. In multiconditioned training, we train acoustic model on various noise conditions and test on seen/unseen noises along with Noise Aware Training (NAT). The results suggest that all front-ends provide performance improvement for seen noise conditions while degrading performance for unseen noise conditions. TDNN-DAE provides the most improvement for seen conditions while giving the most degradation for unseen conditions. We use a method to improve performance of TDNN-DAE in unseen conditions by training it on features enhanced using Vector Taylor Series with Acoustic Masking (VTS-AM) and Spectral Subtraction (SS). We show that these enhancement techniques improve the efficacy of the TDNN-DAE significantly in unseen noise conditions. Overall we observed that multiconditioned training still gives better performance in case of both seen/unseen noise conditions, although the enhanced TDNN-DAE comes closest among all the front-ends to the performance of multiconditioned training.