The Improvement and Implementation of Speech Enhancement Based on Mel frequency Wiener Filtering

Abstract

The main purpose of speech enhancement is to eliminate the noise in noisy speech signal and extract pure speech signal, which has important significance to improve the performance of digital hearing aid. This paper mainly studies the speech enhancement technology in digital hearing aids. Using the improved second order Mel twisted Wiener filtering algorithm, introduced short-time amplitude spectrum of dynamic decision voice activity detection (VAD) algorithm, which solves the part deviation of estimation stationary noise. At the same time, add a priori SNR gain factor based on decomposition of pure noise frame to increase the degree of inhibition, and contain the speech frame is reduced the extent of the suppression. Update the SNR prediction and low SNR ratio and Wiener filtering gain coefficient, automatic gain control (AGC) effect is obvious. The experimental results show that the output SNR of the processed signal is obviously improved, and the speech intelligibility is good and the quality is high. Significantly improve the recognition ability of weak signal. Introduction In the speech enhancement algorithm, the method of adding the Mel frequency domain processing method, which can make the speech more in line with the human ear hearing. The basic theory of Wiener filtering is on the assumption that the input current filter for useful signal and noise, both of which are generalized stationary process, and knowing their second order statistics, Wiener according to the minimum mean square error criterion obtained the parameters of the linear filter, the obtained parameters are designed to filter known as Wiener filters. Wiener filtering is applied to the need to separate the signal from the noise is the whole signal (waveform), rather than its several components. Input hypothesis of the Wiener filter for random signal containing noise, the difference between the desired output and actual output error for the square error, error is small, the effect of filtering noise better. In order to minimize the mean square error, the key is to find the impulse response. If it is able to satisfy the Wiener Hof equation, the filtering effect can be achieved. Speech enhancement based on Wiener filtering, although the quality of the processed sound be significantly improved, but it retains too much background noise, so that the output SNR decreased. The key is to estimate the accuracy of the noise spectrum. So in this paper, this improved the second order Mel twisted Wiener filtering algorithm and the design basis is two stages signal processing, each stage were Mel warped Wiener filtering and processing, output of the first stage as the input signal of the second stage, and after join based on a priori SNR gain factor, for each frame signal of coefficient of adjustment. A fast noise processing method is designed to improve the accuracy of the algorithm. And the robustness of the algorithm is also improved. Mel frequency Wiener filtering basic model In the application of speech enhancement, the input signal y(n) is a band of noise signal, which can be expressed as: ( ) ( ) ( ) y n x n n n = + International Conference on Applied Science and Engineering Innovation (ASEI 2015) © 2015. The authors Published by Atlantis Press 1814 The x(n) is the pure speech signal, n(n) is the noise signal, The purpose of Wiener filter is through filtering the input signal, produce the estimation of pure speech signal x(n). We introduce the derivation process of frequency-domain Wiener filter. In speech recognition and speaker recognition, the common speech feature is based on Mel frequency cepstral coefficients (Mel frequency cepstrum coefficient, MFCC). Due to the MFCC parameters is the human ear auditory perceptual characteristics and speech generation mechanism combination, so the current most speech recognition systems widely used this feature. Here we introduce the derivation process of the Wiener filter in the frequency domain. : Due to the characteristics of human ears, ear perception for these pure tone frequency are nonlinear, we usually use the Mel scale to show, the relationship between the scale and the actual linear frequency can be approximated as: 10 ( ) 2595 log (1 / 700) Mel f f = × + The calculation formula of triangular window function for the Mel filter: 0 ( 1) ( 1) ( 1) ( ) ( ) ( 1) ( ) ( 1) ( ) ( 1) ( 1) ( ) 0 ( 1) k f m k f m f m k f m f m f m W k f m k f m k f m f m f m k f m +  Among them, m is the Mel frequency, k is digital linear frequency. After the calculation of Wiener filter, Wiener filter coefficients can be changed Mel domain: