Confidence weighting missing feature approach for robust speech recognition

Missing feature theory (MFT) has been proposed as a solution for robust speech recognition. It improves robustness of speech recognition systems by either ignoring or compensating the unreliable components of feature vectors corrupted mainly by band-limited background noise. Since the local corruption often occurs in the frequency domain and it is smeared by the Discrete Cosine Transform (DCT) used to obtain cepstral features, algorithms utilizing the missing feature theory are usually restricted to spectral features. In many cases cepstral features might be preferable. In this paper, we propose a new missing feature approach (CWMFA) based on confidence analysis of feature vector and successfully apply it on cepstral features. In the new approach, probabilities of feature vector components are weighted with its confidence in logarithmic domain. Experimental results show that the proposed approach can manifestly improve robustness of speech recognition systems.