基于多窗谱相关加权语音增强

Correlation Weighting Enhancement of Speech Based on Multitaper Spectrum

与传统的周期谱图相比,多窗谱分析方法是一种低方差、高分辨率的谱分析方法,尤其适合非线性系统中强噪声背景下弱信号、时频演变信号的分析。基于多窗谱估计这种优点,提出多窗谱相关加权语音增强方法,先对噪声与含噪信号比(NNSR)进行估计,用基于NNSR的幅度谱减实现预语音增强,再用相关加权规则获得最终的增强语音。通过客观和主观测试表明,在相同的实验条件下,多窗谱相关加权算法能更好地抑制背景噪声和音乐噪声,同时也较好地保持了语音的可懂度和自然度。

Compared with the traditional periodogram,the Multitaper method of spectral analysis provides a means for spectral estimation with low variance and high resolution.It is particularly well-suitd for the diagnosis analysis of weak signals with a time-depended amplitude and frequency against a high-noise background.Because of the low variance feature of the multitaper spectrum,an algorithm of correlation weighting enhancement of speech based on multitaper spectrum is presented.The noise spectrum and the Noise to Noisy Signal Ratio（NNSR） are estimated based on the multitaper spectrum of the noisy signal,and the pre-enhanced speech is obtained by the spectral amplitude subtraction method,whose gain is a function of NNSR.The final enhanced speech is obtained by correlation weighting rule,and subjective and objective tests are made on this algorithm.The results show that this algorithm is very effective to reduce the background noise and music noise,moreover,ensures the intelligibility and naturalness of speech.