Study Of Speech Enhancement Technology Under Low SNR

This thesis can be divided into two parts .One is the endpoint detection of noisy speech signals. The other is speech enhancement in the low SNR circumstances. The latter is the main part of this thesis. The first part discusses about the endpoint detection. The endpoint detection is basic technology of the speech signal processing. It plays an important role in speech recognition, speech enhancement and speech coding. The previous methods either require the statistical information of the speech and the noise which make it difficult to perform in time such as the method based on HMM model, or can not perform well in low SNR circumstance such as the method based on short energy. In this thesis, we bring forth a new method based on intermittent chaos of Duffing oscillator. The experimental results show that this method performs better than that based on short energy in low SNR. The second part concentrates on speech enhancement. Speech enhancement is one of the most important parts of the speech signal processing. It also plays an important role in speech recognition and speech coding. This thesis firstly introduce the different characteristics between speech, hearing and noise, discuss various of speech enhancement methods, then give a solution of "musical noise" and put forward an improved spectral subtraction method. This method is Harmonic enhancement based on periodicity of speech. A number of experiments showed that this method has better performance than others. In addition, according to the theory of wavelet transform and the idea of spectral subtraction, a modified spectral subtraction method on wavelet domain is presented in this thesis. Both of the two methods can improve the SNR greatly, and remove the music residual noise effectively. The result shows the secondly method performs better.