Research On Speech Enhancement In Non-stationary Noise Environments

In daily life, speech makes people communicate with each other efficiently. But, noise often interferes the speech signal, it reduces the quality of speech, and causes bad influences when people are communicating with each other. Usually, noise has different characteristics and is random, it is impossible to eliminate noise completely, especially in non-stationary noise environment. So, the technology of speech enhancement is to eliminate noise and improve the quality of speech.Starting from studying the traditional speech enhancement algorithms, we analyzed their advantages and disadvantages. On the basis of the auditory masking effect, we proposed some improvements on a single channel speech enhancement algorithm. When we are in the environment that speech and noise come into our ears at the same time, we enhanced the speech signal by using an algorithm which bases on perceptual distortion measure. The main work and innovations are as follows:1. The algorithm of speech enhancement based on noise removal1) In order to solve the problem that speech signal is easily distorted when we are enhancing speech signal, in the thesis, we introduced a low distortion speech estimator. However, the speech estimator has high computational complexity, when SNR is low, the estimator cannot reduce noise accurately. To solve this problem, we proposed an over-subtraction factor, it reduced the computational complexity. After that, the estimator can adjust the gain function according to SNR in time.2) In order to obtain a speech enhancement system which has high performance, it is necessary to estimate the noise spectrum accurately. In this thesis, we introduced an unbiased noise power estimation measure with low complexity and low tracking delay. It can update the noise spectrum of every frame in time, and it also can estimate the spectrum of the non-stationary noise accurately. The experimental results show that it has better performance than the traditional algorithms.3) When pre-estimating the speech signal, we choose an algorithm which is based on Minimum Mean-Square Error Log-Spectral Amplitude. The pre-estimating speech signal conforms to the human hearing characteristics. The experimental results show that it has better performance than the traditional algorithms when combines with the above mentioned noise spectrum algorithm.2. The algorithm of speech enhancement based on auditory perceptionWhen we are in the environment that speech signal and noise come into our ears at the same time, the speech signal should be pre-processed, in this thesis we presented an speech enhancement algorithm based on perceptual distortion measure. By simulating the human auditory system, the energy of the speech signal will be redistributed, so the distortion which we can perceive will be minimized. In the speech enhancement system based on perceptual distortion measure, we can control the delay of the algorithm in the different application scenarios, we can obtain different enhanced efficiency. The experimental results show that it can achieve speech enhancement in noise environment.