连续语音识别前端鲁棒性研究  

Front-end Robust Study for Continuous Speech Recognition

作　　者：胡丹[1] 曾庆宁[1] 龙超[1] 黄桂敏[1] 

机构地区：[1]桂林电子科技大学信息与通信学院,广西桂林541004

出　　处：《电视技术》2015年第24期43-46,58,共5页Video Engineering

基　　金：国家自然科学基金项目(61461011;41201479);广西自然科学基金项目(AA053232;BA118273);桂林电子科技大学研究生科研创新项目(GDYCSZ101456)

摘　　要：针对大词汇量连续语音识别中识别率不高的问题,提出了将语音增强级联在识别系统前端,在语音增强中将谱减法和对数最小均方误差算法(logmmse)与用于噪声估计的最小控制递归平均算法(imcra)相结合。识别系统使用Mel频率倒谱系数(MFCC)提取特征,用隐马尔科夫模型(HMM)训练与识别。实验结果表明,该方法最高能使单词识别率提高38.9%,使句子正确率提高21.8%。该方法用于大词汇量连续语音识别是可行有效的。For the accuracy rate of large vocabulary continuous speech recognition is low, using a speech enhance system before recognition is proposed. In this system, spectral subtraction and logarithmic minimum mean square error （logmmse） is combined with improved minimum controlled reeursive average algorithm （imcra） which for noise estimation. In recognition system, Mel Frequency Cepstral Coefficients （MFCC） is used to extract features and Hidden Markov Model （ HMM ） is used to training and recognition. Experimental results show that word recognition rate increased by 38.9% and sentence accuracy rate increase by 21.8% due to the use of the method proposed in this paper. The method used for large vocabulary continuous speech recognition is feasible and effective.

关 键 词：连续语音识别 语音增强 HMM imcra 句子正确率 

分 类 号：TN912.34[电子电信—通信与信息系统]