基于正弦模型的语音识别时频特征

为改善语音识别系统的性能，采用时频分布参数来描述语音特征。由于时频分布参数考虑到语音信号内在的非平稳特性，因此能够更准确地描述语音信号的时频特性。对基于正弦模型的多种时频参数(能量谱和幅度加权瞬时频谱)进行了比较，并在基于隐马尔可夫模型的连接词语音识别系统中进行了实验仿真。结果表明，单独采用时频分布参数作为ASR的前端特征并不能改善识别率；而采用标准ASR特征和能量谱时频特征的联合前端特征，可以有效地改善语音识别系统的识别效果。

Time-frequency Features for Speech Recognition Based on Sinusoidal Speech Model

The use of time-frequency distribution as features for automatic speech recognition (ASR) is intended to improve the performance of ASR. Because of the inherent non-stationary characteristic of speech signal, time-frequency distribution can describe the signal more precisely. Time-frequency features,such as energy spectrum and amplitude weighted short-time average of the instantaneous frequency,are compared based on the sinusoidal speech model, and some experiments are carried out in connected digits recognition system based on Hidden Markov Models. The results indicate that time-frequency features as front-end can't improve the performance of ASR solely,while the feature which combines standard ASR features and time-frequency features based on energy spectrum can improve the performance of ASR system effectively.