Research On Speech Enhancement Based On Noise Spectrum Estimation And Signal To Noise Ratio Constraint

Speech signal is the most common communication style for human beings. However, the speech does inevitably suffer from noise of the surrounding environment. Therefore, speech enhancement is necessary required before post treatment of speech signal corrupted by noise. It is significant for seeking an effective algorithm to achieve higher speech quality. For single channel speech enhancement algorithms, speech enhancement consists of two stages, that is, noise spectrum estimation as well as speech denoising using the estimated noise spectrum. The accuracy of the noise spectrum estimation has positive impact on denoising effect. The speech spectrum gain function also has influence on the speech quality. This thesis focuses on speech enhancement technology for single channel condition. The main research work is summarized as follows:1) We proposed a noise spectrum estimation method based on the IMCRA (Improved Minima Controlled Recursive Averaging) algorithm. The IMCRA algorith--m is one of the most popular noise spectrum estimation methods. However, it does not update the noise spectrum during periods of speech presence. Therefore, it cannot track noise variations in time and cannot estimate the noise spectrum effectively. In this thesis the proposed IMCRA based noise spectrum estimation method updates the noise spectrum of the current frame using neighboring noise spectrum of the previous frame during speech presence. The experimental results indicate that the speech quality of the enhanced outperforms that of the IMCRA algorithm.2) We proposed a statistic based speech enhancement algorithm under the constraint of the signal-to-noise ratio (SNR). Statistic based speech enhancement methods derive the gain function through minimizing the mean square error (MMSE) of the estimated speech spectrum and the real clean speech spectrum. The enhanced speech signal can be obtained by multiplying the gain function and the noisy amplitude spectrum. However, MMSE-STSA pays no attention to positive or negative differences between the clean and estimated spectra, then more noise would be remained in the enhanced signal. In this thesis, we propose a method to adjust the short time spectrum estimation gain function based on MMSE by adding constraints on the a prior SNR and the posteriori SNR, that is, the value of the gain function is constraint by the value of the SNRs, and weaken the noise. The experimental results show that the speech quality is higher than that of the conventional MMSE algorithm.