Research Of Several Problems On Speech Enhancement Systems

Encoding and decoding, transmission and speech recognition technologies of pure speech have been well developed in the speech communication system. However, in the circumstances with background noise and channel noise, the performance of signal process system will degrade dramatically, and then impact on speech signal quality. The research on the speech enhancement has lasted for many years, and there are some achievements on better performance in the case of stationary and slow time-varying noise environment. However, in the complicated environment with fast time-varying and non-stationary noise environment, there are many deficiencies by using the existing noise reduction technologies, and it may cause big influence on the speech quality and intelligibility. The performance of enhanced speech in the signal processing system is worse than that of pure speech. Based on the above issues, this thesis mainly focuses on the front-end process of speech in the speech enhancement and related issues.1. Based on the discussion of speech enhancement algorithms, we summarized and classified them into two types, single-channel speech enhancement algorithm and multi-channel speech enhancement algorithm. These two algorithms are described in this thesis, and they are the basic theory of further research.2. The estimation of the a priori signal-to-noise (SNR) is a crucial part of speech enhancement algorithms. In order to solve the delay issue in the existing a priori SNR estimation algorithms and trace the speech signal in the noise environment, a new a priori SNR estimation algorithm is proposed. It takes the influence of both a priori SNR and posterior SNR to calculate smoothing factor, and can solve the problem of jitter caused by posterior SNR and the problem of delay caused by a priori SNR. The correlation of inter-frame and intra-frame is also considered in the proposed algorithm. Smoothing factor is calculated each frame, and make the noisy speech be handled with different smoothing. At last, the updated a priori SNR is applied to the speech enhancement system, and simulation is made to evaluate the performance in the modified speech enhancement system. The simulation results show that the proposed algorithm improved the performance of speech enhancement system and is better in the non-stationary noise environment. The proposed algorithm can be widely used in the speech enhancement system based on short-time spectrum estimation.3. The frequency domain speech enhancement algorithms is one of the most widely used algorithms. The existing frequency domain speech enhancement algorithms do not consider the speech signal, and lead to short-time spectral peaks, which is caused by the smoothing algorithm in frequency domain. The proposed algorithm transforms the related parameters of frequency domain speech enhancement algorithm to the cepstral domain first. Then make different smoothing to the envelope, fine characteristic and noise in order to restrain the spectral peaks, make compensation on the frequency domain algorithm and then restrain the noise while protecting the speech characteristics. The proposed algorithm can be widely applied to the existing frequency domain speech enhancement algorithms and estimation of the a priori SNR, and can effectively decrease the musical noise and improve the performance of speech enhancement system.4. Voice activity detection (VAD) is an important enabling technology for a variety of speech-based applications. Considering that VAD based on direct-decision likelihood test cannot well catch the variation of noise, that results in error decisions, we propose a new VAD algorithm based on the adaptive threshold likelihood ratio test. In the proposed algorithm, different decision thresholds are set according to the SNR each frame; then decisions are made using the adaptive threshold. Simulation results show that the proposed algorithm improved the performance of VAD algorithm, and solved the problem of error decisions which exist in the VAD based on direct-decision likelihood test.