Research On Application Of Pitch Extraction Method Based On Speech Enhancement To Speech Recognition

The tone features play an important role in Mandarin speech recognition system. Tone is acoustically characterized by the pitch contour. However, in real environments, accurate and robust pitch extraction method has been a difficult problem. So how to extract robust pitch has become one of the important research fields of mandarin speech recognition. Also tone plays an important role in disambiguating confusable words. But traditional single-stream HMM(Hidden Markov Model) method hasn't explored the tone features effectively. So how to model tone feature effectively, also has become one of important and hot fields of mandarin speech recognition.The main content of this thesis are shown as follows:1. Construct a robust and effective pitch extraction algorithm. In pratical noise environment, traditional pitch extraction algorithm has shown bad performance. In this condition, we choose speech enhancments method as pre-process unit of pitch extaction algorithm. This method can decrease the noise influence on speech spectrum, and also has been proven effective in noisy environments, especially under car environment.2. To solve traditional single-stream HMM method can not explord pitch feature effectively, we choose two-stream HMM method model spectrum features and tone featrues, respectively. In this new model method, each HMM state unit can include two content. The first one is used to model spectrum features, the second one is used to model tone featrues.Experimental results show that the proposed method yields relative significant improvement.