摘要
针对噪声环境下语音识别系统性能下降的问题,提出一种基于语音时频相关性的Mel特征矢量聚类补偿算法。该算法首先实现掩码估计,利用纯净语音信号时域和频域的相关性,实现了时频块的有效划分和基于时频块的语音特征聚类。在此基础上,对带噪语音的Mel语谱进行特征补偿。采用HTK工具和TIDIGITS数据库加入不同类别噪声的语音测试结果表明:该算法在不同信噪比条件下,获得了较基于频域相关性聚类特征补偿算法更好的性能。
The performance of automatic speech recognition systems declines dramatically in noisy environments.This paper presents a missing data technology（MDT） feature compensation method based on a time-frequency correlation（TF-CBFC） to improve the speech recognition.This method first estimates the missing data masks and then divides the spectrogram into time-frequency blocks（TFBs）.Clusters of the TFBs are used to compensate for missing data in the Mel spectrogram of noisy speech.Tests with the HTK tools on the TIDIGITS database using different classes of noise show that this method outperforms the MDT feature compensation method based on the spectral relations for various signal to noise ratios（SNRs）.
出处
《清华大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2013年第6期753-756,共4页
Journal of Tsinghua University(Science and Technology)
基金
国家自然科学基金项目(61271309)
