Feature-dependent compensation of coders in speech recognition

Abstract

A solution to the problem of speech recognition with signals corrupted by coders is presented. The coding-decoding distortion is modelled as feature dependent. This model is employed to propose an unsupervised expectation-maximization (EM) estimation algorithm of the coding-decoding distortion that is able to cancel the effect of coders with as few as one adapting utterance. No knowledge about the coder is required. The feature-dependent adaptation can give a word error rate (WER) 21% lower than the feature-independent model. Finally, when compared to the baseline system, the reduction in WER can be as high as 70%.