Research On Noise Estimation In Speech Enhancement

The purpose of speech enhancement is to enhance the target speech, in other words, to reduce noise. Many speech enhancement algorithms have been proposed after decades of research. And for many speech enhancement algorithms, noise estimation is a very critical part.Speech signal is non-stationary but with short-term stability. So noise estimation algorithms generally estimates noise power spectral density through short time spectrum analysis by using mathematical methods such as stochastic process and probability statistics. In this paper, the research is focus on noise estimation of speech in mono situation. Several typical noise estimation algorithms are introduced and simulated. And for which the Improved Minima Controlled Recursive Averaging(IMCRA) algorithm and noise estimation based on Minimum Mean Square Error are improved.Due to the non-stationary of noise signal, the power spectral density of noise changes during speech presence. So it is necessary to track noise rapidly. IMCRA algorithm carries smoothing and minimum search twice comparing to Minima Controlled Recursive Averaging(MCRA) algorithm. IMCRA algorithm searches the minima by taking a one-way search, in the case of increasing noise levels, this method can’t catch the rapid change of noise power, and may result in large tracking delays. In this paper, the minimum search of IMCRA has been improved. For the first time of minimum search, use a two-way search to get a rough speech presence judgment. Then for the second time of minimum search, use a two-way search with a long frame and a short frame at the same time, and combine the search result with the judgment of the first search to yield the final minimum search result. Simulation results show that the improved algorithm can reduce the delay of noise estimation and better estimate the noise.In the noise estimation algorithm based on Minimum Mean Square Error(MMSE), the MMSE estimation of noise periodogram is obtained by using limited maximum likelihood estimation of the a priori SNR. However, under the given a priori SNR estimation, the resulting MMSE estimate exhibits a bias which can be computed analytically. In order to compensate for the bias, a second estimation of the a priori SNR is required. To improve the noise estimation based on MMSE, the original algorithm is modified such that it evolves into a soft speech presence probability instead of a hard based estimator, which automatically makes the estimate unbiased. From experimental results it followed that the modified algorithm generally achieves similar or even better performance as the original MMSE algorithm and with the advantage that no bias compensation is needed and the computational complexity is even lower.