Sound source separation for robot audition using deep learning

Abstract

Noise robust speech recognition is crucial for effective human-machine interaction in real-world environments. Sound source separation (SSS) is one of the most widely used approaches for addressing noise robust speech recognition by extracting a target speaker's speech signal while suppressing simultaneous unintended signals. However, conventional SSS algorithms, such as independent component analysis or nonlinear principal component analysis, are limited in modeling complex projections with scalability. Moreover, conventional systems required designing an independent subsystem for noise reduction (NR) in addition to the SSS. To overcome these issues, we propose a deep neural network (DNN) framework for modeling the separation function (SF) of an SSS system. By training a DNN to predict clean sound features of a target sound from corresponding multichannel deteriorated sound feature inputs, we enable the DNN to model the SF for extracting the target sound without prior knowledge regarding the acoustic properties of the surrounding environment. Moreover, the same DNN is trained to function simultaneously as a NR filter. Our proposed SSS system is evaluated using an isolated word recognition task and a large vocabulary continuous speech recognition task when either nondirectional or directional noise is accumulated in the target speech. Our evaluation results demonstrate that DNN performs noticeably better than the baseline approach, especially when directional noise is accumulated with a low signal-to-noise ratio.