Study On Noise Estimation-based Speech Enhancement Approaches

In real life, because the speakers are often in the noisy environments, the speech signals are unavoidable to be influenced by the noise, which not only makes the speech be hard to perceive and understand, but also reduces the performance of the signal processing systems severely. In order to reduce the influence of background noise on the speech signals, speech enhancement is executed to suppress the background noise and improve the speech quality. Speech enhancement is an important content of the speech signal processing. On the one hand, through carrying out speech enhancement, you can improve the clarity, intelligibility and comfort of speech, and also improve the quality of human auditory perception; on the other hand, speech enhancement is an essential procedure of the speech signal processing systems, in order to reduce the impact of the noise on the speech signal processing systems and improve the performance, the speech signals should be enhanced firstly.The research of speech enhancement stems from the demand for technology in the human life, and has an important practical application value. This dissertation mainly focuses on the speech enhancement based on noise estimation, proposes several new noise estimation and speech enhancement algorithms based on time-frequency correlation and noise classification to overcome the certain limitations of existing algorithms, and validates the effectiveness and universal applicability of our proposed algorithms through theoretical proof and a large number of experiments. The main innovation of this thesis is as follows:1) We propose a noise estimation algorithm based on time-frequency correlation. The correlation between time-frequency units is a significant feature for judging the presence of the speech in the noisy speech. By calculating the correlation function between the time-frequency units, we obtain a rough decision about speech presence, using which we improve the key steps in the conventional improved minima controlled recursive averaging (IMCR.A) algorithm, and propose a new noise estimation algorithm. The experimental results show that our proposed algorithm can estimate the noise more accurately, and significantly reduce the delay of the noise estimation in the case of noise bursts.2) We propose a noise classification method based on the energy distribution in the Bark domain. In the actual environments, there are a great many types of noise, and they affect the speech signals in different ways. To classify various noise and process it according to the noise types can improve the performance of noise estimation and speech enhancement. In this thesis, by analyzing the noise energy distribution characteristics in the Bark domain, we obtain an18-dimensional feature vector. Adopting the support vector machine (SVM) classifier, we propose a noise classification method with high classification accuracy.3) We propose a speech enhancement algorithm based on noise classification. Firstly, we obtain the optimal parametric combinations for the IMCRA under different types of noise by enhancing the noisy speech in the training set. Secondly, we propose an effective method to judge the noise type of noisy speech based on the previous noise classification method, and carry out the noise estimation using the optimal parametric IMCRA according to the judged noise type. Finally, the speech is calculated by the optimally-modified log-spectral amplitude estimator. The results of speech enhancement experiments under different noise conditions show that the proposed algorithm can preserve the speech components better, suppress the background noise more effectively, and improve the overall quality of enhanced speech.4) We propose an improved minimum search method for noise estimation. The minimum search of the noisy speech power spectrum is the foundation of noise estimation. To improve the accuracy of noise estimation and reduce the delay of noise estimation in the case of noise bursts, we propose a parallel search method which adopts two minimum searches with different window lengths to search the minima at the same time. The final results of the minimum search are determined by the results of the two searches and the binary decision about speech presence based on noise classification together. The experimental results show that, for highly non-stationary noise, the proposed minimum search method can effectively reduce the delay of noise estimation and significantly improve the quality of enhanced speech.