基于时频ICA的PMC模型卷积噪声估计方法研究

A study on PMC convolutive noise estimation method based on time/frequency ICA

为提高卷积环境下语音识别系统的鲁棒性,提出了一种基于时/频ICA(independent component analysis)的卷积噪声模型估计方法.所提算法首先使用ICA方法从含噪语音信号中提取纯净语音信号的短时功率谱,然后在MEL滤波器组域内将含噪语音的短时谱减去纯净语音的短时谱,并根据去噪后卷积噪声的短时谱估算其HMM(hidden markov model)模型.在仿真和真实环境下进行了语音识别实验,其识别正确率相比较传统的卷积噪声估计方法分别提升了4.70%和4.75%.实验结果表明,论文所提算法能够实现对卷积噪声的精确估计,并有效提升卷积噪声环境下语音识别系统的性能.

In order to improve robustness of speech recognition system in convolutive environment, a convolutive noise estimation method based on time/frequency ICA （independent component analysis） （TD-ICA） was proposed in the paper. The algorithm firstly separated the short-time spectrum of speech and noise by TI）-ICA algorithm, and then the noise short-time spectrum was acquired by subtracting the estimated clean speech short time spectrum from the noisy speech in the mel-scale filter bank domain. Finally, an HMM （hidden Markov model） of convolutive noise was established based on the noise short-time spectrum. Experiments have been carried out in simulation and real environment, experiential results revealed that the proposed algorithm obtained the relative increasing of 4. 70% and 4. 75% compared with conventional noise estimation method, which validated the accuracy of estimated noise signal and proved that the proposed algorithm could effectively improve recognition ratio in convolutive noise environment.