基于多频带谱减法的抗噪声语音识别研究  

Robust Speech Recognition Based on Multi-band Spectral Subtraction

作　　者：万义龙[1] 张天骐[1] 王志朝[1] 金静[1] 

机构地区：[1]重庆邮电大学信号与信息处理重庆市重点实验室,重庆400065

出　　处：《电视技术》2013年第23期183-187,共5页Video Engineering

基　　金：国家自然科学基金项目(61071196;61102131);教育部新世纪优秀人才支持计划项目(NCET-10-0927);信号与信息处理重庆市市级重点实验室建设项目(CSTC2009CA2003);重庆市杰出青年基金项目(CSTC2011jjjq40002);重庆市自然科学基金项目(CSTC2010BB2398;CSTC2010BB2409;CSTC2010BB2411;CSTC2012JJA40008);重庆市教育委员会科研项目(KJ120525)

摘　　要：为了减少在噪声环境下测试条件与训练条件不匹配导致的语音识别性能下降,提出了一种结合多频带谱减法的抗噪声语音识别系统。首先提取带噪语音的前几帧作为估计的噪声信号,将带噪语音、估计的噪声信号按频率划分M个互不相交的频带,然后根据每个频带内带噪语音与估计噪声信号的信噪比,来确定该频带噪声的谱减参数。语音增强作为前端处理,与语音识别器级连构成抗噪声语音识别系统。通过实验仿真表明,基于多频带谱减法的抗噪声语音识别系统在不同信噪比、不同类型的噪声下,识别性能明显优于基本谱减法。In order to reduce the degradation of the speech recognition accuracy while the testing condition are mismatched with the training condition around noisy environment, one kind of multi-band spectral subtraction is put forward. The estimated noise signals are extracted from the first few frames of the noisy speech, and the noisy speech and estimation of noisy signals by the frequency were divided into non-interacted M frequency bands, and then according to the SNR （signal-to-noise ratio） of noisy speech in each frequency band to determine the band noise spectral subtraction parameters. Speech enhancement as the front-end processing module and concatenate with speech recognizer constitute robust speech recognition system. The results of simulated experiments indicate that the recognition accuracy of multi-band spectral subtraction robust speech recognition system is obviously superior to the basic spectral subtraction in different signal-to-noise ratios and different noise＇ s types.

关 键 词：语音增强 多频带 谱相减 语音识别 隐马尔可夫模型 

分 类 号：TN912.34[电子电信—通信与信息系统]