Non-stationary Noise Estimation Methods In Speech Communication

The existing single channel speech enhancement technologies perform well inthe tracking and suppression of stationary noise. But for the non-stationary noise inreal life, the noise estimation is not so accurate and the performance of noisereduction declines dramatically. As a result, speech enhancement in non-stationarynoise environment has become an important issue for real-world applications.The content of this thesis is mainly focused on the following aspects:Firstly, the basic principle of Hidden Markov Model (HMM) and its applicationfor speech enhancement are reviewed. The HMM applied for speech enhancement iscompared with that for speech recognition. Then the definition and principle ofAuto-Regressive HMM (AR-HMM), which is used in this thesis, areintroduced.Secondly, HMM-based speech enhancement method for non-stationarynoise environment is proposed. The AR-HMM of clean speech and noise signal aretrained (their parameters) using Baum-Welch algorithm respectively and the featuresof HMM are composed of the normalized excitation energy and the Line SpectrumFrequency (LSF) parameters. In the enhancement stage, the normalized critical bandpower spectrum is used as the feature in Gaussian Mixture Model (GMM) to classifythe background noises. Based on the AR-HMM of clean speech and the noise ofcorresponding type, the power spectrums of speech and noise are estimated underminimum mean square error (MMSE) criteria and the noisy speech is enhanced by thegain function of wiener filter. Taking into account the differences between the trainingdata and test data in the non-stationary noise environment, the online adjustmentmethod for the speech and noise models is proposed. The scaling factor of speechenergy is estimated using the iterative Expectation Maximization (EM) algorithm andthe scaling factor of noise energy is estimated using the re-estimation approachsimilar to the training stage. And the initial scaling factor of noise energy is obtainedby Minima-Controlled Recursive Averaging (MCRA) algorithm.Finally, in order to compensate for the change of speech level caused by thesignal collection procedure or the pre-processing module like speech enhancement, acompressed Automatic Level Control (ALC) method operating in the networkequipments is proposed. Real-time speech level is measured using ITU-T P.56standard. According to the difference between the real-time level and the target level,the adaptive and fixed codebook gains in the input speech bit-stream are jointlymodified to (control) adjust the speech level to the comfortable range of listening. The evaluation of the proposed speech enhancement method is performed underthe standard of International Telecommunication Union, TelecommunicationStandardization Sector (ITU-T) G.160. The test results show that, comparing with thereference methods, the proposed speech enhancement method performs well innon-stationary noise environments. Larger noise reduction and shorter convergencetime are achieved by the proposed method. The active level bias of the proposed ALCmethod in the compressed domain is below0.5dB, it has better objective speechquality than the reference method.