A new feature set for masking-based monaural speech separation

摘要

We propose a new feature based on a gammatone filter bank for improving monaural speech separation using neural networks. This new feature encodes not only the local information of cochleagram, and spectrotemporal context, similar to previous approaches, but also captures time-frequency dynamics in the spectrotemporal context using an image processing technique. Speech separation was achieved by computing optimal time-frequency masks using two types of neural networks (DNN and LSTM) to determine the interactions between feature and training model properties. The performance of our feature was evaluated in a variety of simulated environments having different non-stationary noises and reverberation times and quantified using three objective measures. Experimental results show that the proposed monaural feature set improves the objective speech intelligibility, speech quality and signal-to-noise ratio compared to prior feature sets in noisy and reverberant environments with particular benefit in speech intelligibility.