Speech Enhancement Approaches Under Complex Conditions

Speech is one of the most main ways for communication, but the speech signal is often inevitably polluted by noise. Noisy speech will make people's listening experience becomes worse, and even hard to capture the information carried by the speech. In Speech code, Speech Recognition, Speaker Recognition systems, noise can seriously affect the results of the applications. Therefore, the enhancement of speech is necessary.At present, the effect of existed speech enhancement methods can't meet people's requirements, the main problems are as follows: there is a residual noise in the enhanced speech; algorithms are not universal, can only handle certain types of noises; the computational complexity is high, and so on. To solve these problems, this thesis research speech enhancement algorithms under complex conditions. The goal is to make the algorithm to adapt to a variety of noise types and complex environment that changing over time, and to improve the enhance results, reduce the residual noise, and cut down the algorithm's computational complexity. The main works and contributions are:(1) A comprehensive summary and analysis of existed speech enhancement algorithms is given. The Single-channel speech enhancement algorithm are divided into five classes, which are methods based liner filter, methods based on short-time spectral estimation, methods based on speech model, methods based on hearing masking effect and new methods for speech enhancement. Those methods are compared in characteristic, defect and computation complexity. Dual-channel and multi-channel speech enhancement algorithms are also introduced briefly, and the development direction of speech enhancement is discussed.(2)A speech enhancement algorithm of OM-LSA (Optimally Modified Log-Spectral Amplitude Estimator) incorporating wavelet thresholding is proposed and implemented; experiments verify the effectiveness of the algorithm. To reduce the residual noise in OM-LSA algorithm, wavelet thresholding algorithm is introduced, and proposed a method to estimate the residual noise. The original noise variance is calculated by the wavelet decomposition of original noisy speech, combined with the gain function under speech absence in OM-LSA. The detailed process of the proposed algorithm is given, and is achieved. The algorithm is tested using a standard database, and the results show that the proposed algorithm is effective in the reduction of the residual noise, and there is greater improvement in sSNR (segmental Signal-to-Noise Ratio) and LSD(Log-Spectral Distortion) indicators.(3) Speech enhancement algorithm syncretized speech separation based on HMM (Hidden Markov Model) is proposed and implemented; experiments verify the effectiveness of the algorithm. There are two existed speech enhancement algorithms based on HMM which are MAP (Maximum A Posteriori) estimator used under stationary noise condition and MMSE (Minimum Mean-Square Error) estimator used under non-stationary noise condition. Both algorithms have high computational complexity, and the former can't handle non-stationary noise. In response to these shortcomings, with the speech separation technology as reference, speech enhancement algorithm syncretized speech separation based on HMM is designed. First, mix state sequence of noisy speech is decoded under the speech model and noise model; second, the speech is estimated by speech separation method. The proposed algorithm is tested with abundant speech sentences under different noise condition or different SNR. The results show that the proposed algorithm can effectively remove the stationary noise and non-stationary noise, improve the PESQτPerceptual Evaluation of Speech Qualityυscore and the algorithm time is under control.