Exploiting models intrinsic robustness for noisy speech recognition

We propose in this paper an original approach to build masks in the framework of missing data recognition. The proposed soft masks are estimated from the models themselves, and not from the test signal as it is usually the case. They represent the intrinsic robustness of model’s log-spectral coefficients. The method is validated with cepstral models, on two synthetic and two real-life noises, at different signal-to-noise ratios. We further discuss how such masks can be combinedwith other signal-based masks and noise compensation techniques.