Robust Speech Recognition Based on Vector Taylor Series

The vector Taylor series(VTS)expansion is an effective approach to noise robust speech recognition.How-ever,in the log-spectral domain,there exist the strong correlations among the different channels of Mel filter bank and thus it is difficult to estimate the noise variance from noisy speech proposes.A feature compensation algorithm in the cepstral domain based on vector Taylor series was proposed.In this algorithm,the distribution of speech cepstral features was represented by a Gaussian mixture model(GMM),and the mean and variance of noise were estimated from noisy speech by the VTS approximation.The experimental results show that the proposed algorithm can signifi-cantly improve the performance of speech recognition system,and outperforms the VTS-based feature compensation method in the log-spectral domain.