一种基于改进多带谱减法的低信噪比语音识别方法  

Low SNR Speech Recognition Method Based on Improved Multiband Spectral Subtraction

作　　者：陈志博 王思雨 刘远泽 王国栋[1] 耿卫东[1] Chen Zhibo;Wang Siyu;Liu Yuanze;Wang Guodong;Geng Weidong(Key Laboratory of Photo-electronics Thin Film Devices and Technique of Tianjin,Institute of Photo-electronics Thin Film Devices and Technique,Nankai University,Tianjin 300350,China)

机构地区：[1]南开大学光电子薄膜器件与技术研究所,天津市光电子薄膜器件与技术重点实验室,天津300350

出　　处：《南开大学学报（自然科学版）》2022年第6期58-61,共4页Acta Scientiarum Naturalium Universitatis Nankaiensis

基　　金：国家自然科学基金(61401237);中央高校基本科研业务费专项资金。

摘　　要：提出了一种基于相位补偿改进的低信噪比多带谱减算法,构造了相位补偿函数对带噪语音相位谱进行补偿,使其更接近纯净语音相位,提升了多带谱减法在低信噪比下的性能.在Matlab仿真平台上就输出信噪比、PESQ和语谱图进行评估,所提算法均优于多带谱减算法,其中在-10 dB的极低信噪比下,该算法在PESQ上比原算法提高了26.4%.An improved low signal-to-noise ratio(SNR) multiband spectral subtraction algorithm is proposed based on phase compensation, constructed a phase compensation function to compensate the phase spectrum of noisy speech in order to make it closer to the phase spectrum of clean speech, and improved the performance of multiband spectral subtraction in low SNR. On the Matlab platform, the proposed algorithm is evaluated and has better performance than the multiband spectral subtraction algorithm in the output SNR, perceptual evaluation of speech quality(PESQ) and spectrogram. In the extremely low SNR of-10 dB, the proposed algorithm is 26.4% higher than the original algorithm in PESQ.

关 键 词：多带谱减 相位补偿 低信噪比 

分 类 号：TN912.35[电子电信—通信与信息系统]