Research On Speech Enhancement Method At Low Signal-to-noise Ratio

As an important method to restrain noise included in the speech, speech enhancement plays a great role in speech recognition and intercommunion between human and machine. In low Signal-to-Noise Ratio (SNR) condition, speech enhancement plays an even greater role in efficient speech recognition. For speech recognition in heavy noise environment, more efficient speech enhancement method is to be put forward.According to the VoIP recognition requirement of the"Telephone Network Security"project, a subtopic of the"Telecommunication Anti-attack System"of the National High-Tech Research and Development Program of China (863 Program), this dissertation stresses the research on speech enhancement method, to restrain the noise and improve the speech quality, as a preceding guarantee for speech recognition. Main work and contributions of this dissertation are outlined as follows:Firstly, to minimize the remaining noise of spectral subtraction, an improved noise spectral estimation algorithm is proposed. Unlike common spectral subtractions which estimate noise amplitude spectral based on average method, the improved spectral subtraction estimates noise according to the distribution character of noise amplitude spectral. Therefore, the improved algorithm is able to subtract noise in a high precision. The emulational experiment in low SNR condition shows that the performance of this improved algorithm is much better than that of other spectral subtractions.Secondly, for the exact estimation of the speech in low SNR condition, an improved speech estimation algorithm is derived. Presently, speech is assumed to obey to Gaussian Distribution as well as noise in statistical model based speech estimation algorithm, just according to central limit theorem. But the central limit theorem is not applicable for short-time signal. Research on this problem indicates that the speech after DCT is close to Laplacian Distribution. Then, based on the different distribution of speech and noise, this algorithm exactly estimates the speech by maximum of a posterior density (MAP). In low SNR condition, this algorithm estimates the speech more accurately and leads to less distortion than the one with supposition of Gaussian Distribution does.Last but not least important, a new speech enhancement method based on Independent Component Analysis(ICA) is designed and analyzed, on the foundation of the two algorithms presented above. This method takes the enhanced signals with the two algorithms above as input observation signals and separates the speech and noise by ICA. The experiment of specific person's isolated speech recognition indicates that this method is able to heighten the recognition ratio in a large scale in low SNR condition.