Asr-driven Binary Mask Estimation for Robust Automatic Speech Recognition

Additive noise has long been an issue for robust automatic speech recognition (ASR) systems. One approach to noise robustness is the removal of noise information through segregation by binary time-frequency masks; each time-frequency unit in a spectro-temporal representation of the speech signal is labeled either noise-dominant or signal-dominant. The noise-dominant units are masked and their energy is removed from the signal. The ideal binary mask, computed given oracle information regarding the speech and noise sources, has been shown to provide significant improvements in speech intelligibility for humans. In this work, we investigate both methods of incorporating binary masks in ASR and methods for estimating the binary mask. While applying binary masks to separation tasks for humans has been straightforward, the incorporation of binary masks in ASR has proved more difficult. The field of missing data ASR proposes methods to compensate for the effects of binary masks on cepstral feature calculation. We demonstrate, contrary to previous work, the direct use of the ideal binary mask performs at least as well as several missing data techniques when the acoustic features have been variance normalized. Typical methods for binary mask estimation focus on low level acoustic features and little work attempts to incorporate higher level linguistic information in the estimation process. We propose an alternative masking criterion that forces the use of higher level information called the ASR-driven binary mask. The mask is defined by force aligning the true