特定人汉语数码语音抗噪识别方法

An Anti-noise Algorithm for Speaker-Dependent Mandarin Vocal Numbers Recognition

本文提出一种连续隐马尔可夫模型(CHMM)和人工神经网络(ANN)相结合的鲁棒性识别方法，用于噪声环境下特定人数码语音识别。该方法以CHMM的输出作为系统的识别矢量，利用人工神经网络的模式分类和自学习功能，从识别矢量空间中提取语音预识别矢量，再由识别矢量对预识别结果进行识别输出。实验证明，这种基于CHMM/ANN的数码语音识别方法明显地提高了系统的噪声鲁棒性，适用于中小词表语音识别系统。

For the recognition of speaker-dependent mandarin vocal numbers under noisy environments，a robust algorithm combining continuous Hidden Markov Models(CHMM) with artificial neural networks(ANN) is proposed. According to this algorithm，the output from CHMM is taken as the feature vector for recognition and the pre-recognition vector is extracted from the recognition vector space by the ANN，afterwards，the pre-recognition result is further processed to accomplish the final recognition. Experiments prove that the algorithm based on CHMM/ANN structure can greatly improve the noise robustness of recognition systems. It is beneficial for the application to medium or small -scale vocabulary systems.