Large Vocabulary Mandarin Continuous Speech Recognition under Noisy Environment

Noise environment and natural spoken speech, is still a challenging issue for speech recognition. In this paper, study on this field is explored on Mandarin speech, from aspects of signal processing, acoustic model, language model, decoding algorithm, and post processing. The two-phase mel-warped wiener filter algorithm is improved for obtaining noise-robust feature. Segmentation algorithm and gender determination algorithm is proposed based on phone decoder. MMIE training algorithm and noise-adding training technology is exploited. One-pass search algorithm is adopted based on cross-word acoustic model, compact search space is constructed which can accurately describe the acoustic context. New fusing technology of multiple sub-systems is proposed which adopt different acoustic feature. Experiments on three testing sets of Mandarin continuous speech show that, the new system obtains relative error reduction of 39.28% in contrast to the baseline system. And the performance of multiple algorithm modules has also been analyzed in detail.