Compensation of nonlinear distortions in speech for automatic recognition

This paper addresses improvement of automatic transcription of speech distorted already during recording or by consequent processing. We focus on distortions that cannot be represented by most often used models, that is, as an additive noise or a linear convolutive channel distortion. We consider a) signals distorted through overgained microphone preamplifier and b) recordings exhibiting unnatural spectral sparseness, caused by application of excessive denoising or low-bit-rate compression. We demonstrate that these distortions deteriorate ASR accuracy significantly. To compensate, we propose to employ a combination of two general robust speech recognition techniques: a front-end feature normalization method and a channel/speaker adaptation technique. We present a significant improvement of transcription accuracy in the case of lectures distorted during recording, compressed broadcast data and utterances recorded with an inappropriately applied denoising.