基于多时间尺度复合子带噪声语音识别新方法

本文根据多时间尺度分析与子带方法,提出了一种多时间尺度复合子带的噪声环境下语音识别新方法.新方法在不同的时间尺度下分别进行子带特征提取和全带特征提取,并分别进行识别,然后在识别概率层相结合得到最终识别结果.本方法兼有多时间尺度方法和子带方法的抗噪性能.此外,进一步引入频谱差分方法提高语音特征的抗噪性能.对E-SET在NoiseX92下白噪声的识别实验表明,新方法具有良好的抗噪性能.

A New Method for Noisy Speech Recognition Based on Multiple Timescale and Hybrid Sub-Band

We present a new method for noisy speech recognition,which combines multiple timescale analysis with sub-band method.In the new method,sub-band features and full band features are extracted in different timescales,then combined at probabilistic layer.The method has robustness of both multiple timescale method and sub band method.Differential method is also introduced to further enhance the robustness of the features.Evaluated by a task on E-SET database under white noise from NoiseX92 noise database,the new method has higher recognition accuracy than conventional multiple timescale method on both noisy speech and clean speech.