Speech Enhancement Based On HMM In Mel-Frequency Spectral Domain

The traditional single channel speech enhancement methods have made a lot of progress, but the enhanced speech obtained by these methods often cause speech distortion or remain ?music noise‘ and other issues when the non-stationary noise is present in real life. Therefore, how to get a better enhanced speech in non-stationary noise environments becomes a topic which needs to be solved in practical applications. In this thesis, a complete set of speech enhancement method using the Mel-frequency spectral domain hidden Markov model(MFS-HMM) is proposed.The contribution of this thesis is composed of the following three parts:First of all, based on the existing MFS-HMM speech enhancement method, an improved speech enhancement algorithm based on MFS-HMM is proposed. This method is generally considered as a weighted-sum filtering of the noisy speech. In the improved algorithm, both HMMs are parall trained in Mel-frequency spectral domain and in log-magnitude domain in order to solve the spectral distortion problem caused by inaccurate estimation of the filter. Then, the vector Taylor series(VTS) is adopted to estimate the parameters of HMM for noisy speech so that the HMM of noisy speech fits the noisy speech better and the weighted-sum filters are more suitable for noisy speech. The parall training of HMM and the usage of VTS make the background noise suppressed effectively and the subjective and objective qualities of the enhanced speech are improved greatly.Secondly, in order to balance the energy mismatch between the training data and test data, an energy adjustment method is proposed for MFS-HMM based speech enhancement. In this method, the online estimation of the log-energy adjustment factor for clean speech and noise is obtained by the iterative expectation maximization(EM) method, and the parameters of HMMs of clean speech and noise are modified online, respectively. This makes the energy of training data match to the test data better and the effect of energy mismatch on the enhanced speech is reduced efficiently. The subjective and objective qualities of the enhanced speech are obviously improved.Finally, the proposed speech enhancement algorithm is applied into ITU-T G.718 speech coder at 12 kb/s. The Objective and subjective quality test results show that the proposed method is superior to the speech enhancement method embedded in G.718 encoder, and can obtain a high quality for the output speech in noise environment.