Single-Channel Speech Enhancement Based On Noise Amplitude Spectral Estimation

The face-to-face or transferred via electronic devices voice communication is indispensable in the human daily life and work. However, in real life, speech signal is always inevitably affected by many kinds of noises owing to the speaker is usually in a variety of noisy environments. The existence of noises will not only make the speech signal to be difficult to perceive or understood, but also can degrade the performances of the speech processing and transforming systems largely that work in the noisy environment. In order to reduce the influence that these background noises take to the speech signal, the speech enhancement technology is implied to restrain the background noises and improve the quality of speech intelligibility. The speech enhancement recovers the clean speech from the noisy signal as much as possible with different theories and methods, with the purpose of improving the quality and intelligibility of the speech. According to the different methods of speech signal processing, the speech enhancement can be divided into the time-domain and transform-domain methods. Being able to distinguish more prominently between the characteristics'differences of clean speech and noise signal, and favor more to the elimination of the background noise, so the speech enhancement algorithms based on different transform domain have been the Chinese and foreign scholars'researching focuses.The thesis is majored on the single-channel speech enhancement technology based on the estimation of the noise amplitude spectrum in the Discrete Fourier Transform (DFT) domain, and pointing to its core parameter estimation problems the paper provides detailed analysis and effective improvement, and proposes two new noise amplitude spectrum estimation methods and the corresponding speech enhancement algorithms, then analyzes and prove the effectiveness and applicability of the proposed methods. The research content of the paper should briefly summarize as follows:Firstly, on the foundation of the amplitude spectral subtraction, the thesis researches and analyses a speech enhancement method based on Quantity Analysis Based Speech Spectral Recovery (QASSR), and states two defaults of the method:1. There is too much residual noise in the output speech of this algorithm; 2. It assumes that the noise amplitude spectrum is known. Pointing to the first default, we propose an improved QASSR speech enhancement based on speech present probability and the priori SNR estimate algorithm based on Decision-Directed. Then, we have proved the superiority of the improved speech enhancement method through experiments and simulations.Secondly, as to the second default of the QASSR speech enhancement, we researched and analyzed the noise spectrum estimate algorithm based on traditional VAD, and stated the deficiency while dealing with the non-stationary noise. By calculating the relation of power spectrum between amplitude spectrum of noise, based on obtaining the noise power spectrum utilizing the soft-decision method, we proposed an indirect method of noise spectrum amplitude estimation. Besides, by introducing the Bayes risk function, we proposed a new noise amplitude spectrum estimation based on minimum mean square error (MMSE) in the Gaussian statistics model. We have simulated and analyzed the effectiveness of the three noise amplitude spectrum estimation methods through experiments.At the last, we fused the improved QASSR speech enhancement method proposed in this thesis with the noise spectrum estimate methods, and got an integral speech enhancement system based on noise amplitude spectrum estimation. By utilizing the speech enhancement system in the different types of noises and input SNR environments, we have tested the speech enhancing effectiveness, and showed the good performance with the simulating results given by different evaluation rules.