Research Of Speech Enhancement Algorithm Based On Short-time Log Spectral Amplitude Estimation

Speech signal is inevitably corrupted by variety of noise in practical applications, such as speech recognition, speech communication, speech coding and speech synthesis, which will interrupt the processing of speech signal in these conditions. In this case, it is necessary to erase the noise from interrupted speech signal.Among all the speech enhancement methods, Short-time log-spectral amplitude minimum mean-square error(LSA-MMSE) is not only simple but also convenient in real-time processing and efficient in removing the noise. It is widely used to eliminate background noise. However, there is still many residual noise or musical noise in its enhancement speech, especially in the case of low signal-to-noise ratio(SNR), which is easy to make a sense of fatigue. In order to improve the de-noising performance of this method,many research works have been done. This subject is supported by the R&D Project of Science and Technology Innovation Commission of Shenzhen, which comes from Research Institute of Tsinghua University in Shenzhen. The project name is research and development of array digital speech processing technology in intelligent digital television, No.CXZZ20130517113418268. Many studies have been done as follows:(1)Analyzing the gain function of LSA-MMSE to figure out the the estimation of priori signal-to-noise ratios which deeply affect performance of the algorithm. Burg spectrum and speech absent probability(SAP) are used to improve estimation of priori signal-to-noise ratios(SNR) in improved method. Burg spectrum is introduced for noisy speech power spectrum estimation. The SAP is utilized to adjust smoothing parameter. In order to test the improved method, audio acquisition system has been built based on Lab VIEW and many simulation experiments have been done. Simulation results show that the improved algorithm has good ability to de-noising.(2)There are many low frequency noise components in enhancing speech of improved LSA-MMSE algorithm when speech corrupted by Volvo noise and office noise. In order to solve this problem, speech enhancement based on improved LSA-MMSE and EMD(Empirical Mode Decomposition) has been studied. When the noises have been removedmostly by using improved LSA-MMSE in low SNR, the signal is decomposed into several Impirical Mode Functions(IMFs) by EMD. Several IMFs were selected to reconstruct enhancement speech according to the IMF statistics. Many simulation experiments have been done on MATLAB platform. Simulation results show that the speech enhancement based on improved LSA-MMSE and EMD can well balance speech enhancement and voice distortion.For example,the speech enhancement based on improved LSA-MMSE and EMD can achieve0.67 d B segment SNR improvement than improved LSA-MMSE on average and improve2.63% degree of distortion than improved LSA-MMSE on average when speech degraded by Volvo noise.