Research On Speech Enhancement Algorithm In The Tunable Q-factor Wavelet Transform Domain

Speech signal is inevitably interfered by noise from the surrounding complexenvironment and transmission medium in the process of acquisition, transmission andreceiving. As a result of noise, speech quality and intelligibility degenerate. Seriously, itwill impact on the performance of speech communication systems. The object of speechenhancement is to obtain the estimated speech from the noisy speech as far as possible.It has a profound influence on the subsequent speech signal processing. Therefore, theresearch of speech enhancement has extensive applied value.Wavelet transform is a local analysis tool in the time-frequency domain. It has thecharacteristics of multi-resolution, adapts to the time-varying signal processing andconquers the shortcoming of Fourier transform which cannot describe the localcharacteristics of signal. So it is widely used in the field of signal processing. Thetunable Q-factor wavelet transform (TQWT), which the Q-factor is easily specified,provides shift invariance, has the perfect reconstruction property and is modestlyovercomplete for discrete-time signal. Furthermore, it can be well localized speechsignal. Therefore, it has become a research topic.The main research work in this thesis is summarized as follows:1. The thesis summarizes the current situation of speech enhancement algorithms andwavelet analysis. The basic knowledge of speech enhancement, such as thecharacteristics of speech signal and noise, the model of noisy speech signal and themain steps of speech enhancement based on wavelet threshold method, and performanceevaluation are reviewed and the basic theory of wavelet analysis is summarized briefly.The thesis is mainly focused on the theoretical knowledge of the tunable Q-factorwavelet transform and its algorithms.2. Aiming at the limitations of methods on speech enhancement by traditionalthreshold methods in wavelet domain, this thesis proposes an improved Donohothreshold and a new threshold function. Based on the tunable Q-factor wavelettransform, the improved Donoho threshold fully takes into account the adaptivitybetween the different scales and voiced and unvoiced signal. Improved thresholdfunction, which is continuous, can flexibly choose in the range of hard and softthreshold function by adjusting the parameter, and avoid the problem of constantdeviation produced by soft threshold function.3. With the study of speech signals’ characteristics in time-domain and the theory of TQWT，the thesis proposes a new speech enhancement algorithm based on the tunableQ-factor wavelet transform. Firstly, it separates voiced signal and unvoiced signal withzero-crossing ratio and short-time energy. Then improved Donoho threshold value andthreshold function are applied to process wavelet coefficients of voiced signal andunvoiced signal, and estimates the original coefficients from noisy coefficients. Lastly,the inverse transform is used to obtain the original speech signal which the noise isremoved. Comparing with the other current classical algorithms, experimental resultsshow that the modified algorithm not only effectively overcomes the spectrumdistortion in spatial domain, but also achieves better effect than the traditional wavelettransform.