摘要
单通道语音信号在信噪比较大的环境下经过增强后再识别,能表现出较高的识别率。但是在低信噪比环境下,增强后语音信号的识别率急剧下降。针对此种情况,提出了一种用在识别系统前端的语音增强算法,该增强算法将采集到的带噪语音信号先使用对数最小均方误差(Logarithmic Minimum Mean Square Error,Log MMSE)提高其信噪比,然后再利用改进的维纳滤波去除噪声残留并提升语音可懂度,最后用梅尔频率倒谱系数(Mel-Frequency Cepstral Coefficients,MFCC)和隐马尔科夫模型(Hidden Markov Model,HMM)对增强后的语音信号做特征提取并识别。实验分析结果表明,该方法能有效地抑制背景噪声并减少噪声残留,显著提升低信噪比环境下语音识别的准确性。
The accuracy rate of single channel enhanced speech recognition in high SNR environment is acceptable, but not so in low SNR environment. In this case, speech enhancement based on logarithmic minimum mean square error（Log MMSE） algorithm and modified Wiener filter algorithm is presented. Firstly the gathered speech signals＇ SNR is improved by the Log MMSE algorithm. Then using the improved Wiener filter algorithm removes residual noise and improves the signal quality. Finally the enhanced speech is used for recognition by MFCC and HMM algorithms. Experimental results show that the proposed method can effectively remove the background noise and reduce the residual noise, significantly increase the accuracy of the automatic speech recognition in noisy environment.
出处
《声学技术》
CSCD
北大核心
2017年第1期50-56,共7页
Technical Acoustics
基金
国家自然科学基金(61461011)
教育部重点实验室2016年主任基金(CRKL160107)资助项目
