利用语音的频谱空间特征进行汉语抗噪语音识别的方法

Spatial characteristics of speech spectrum for robust speech recognition

抗噪连续语音识别是当前汉语连续语音识别的重要研究领域。采用通过度量连续语音帧之间频谱的稳定性,将连续语音切分成份,再将切分结果(无论时间长短)变换为与时间无关的大小固定的频谱空间特征,通过与模板库进行比较实现语音识别。新的频谱空间特征,与语音时长无关,同时表现出较好的抗噪声能力。在特定人连续语音识别测试系统中,取得了不错的识别效果。

The anti-noise continuous speech recognition is an important research topic of current Chinese continuous speech recognition. In this paper, by measuring the stability of the frequency spectrum between the continuous speech frames, speech signal can be segmented, and then these segmentations（regardless of the length of time） are transformed into time-independent and size-fixed spatial characteristics of speech spectrum. By comparing to speech template, the speech recognition results are obtained. The new spatial characteristics of speech spectrum are independent of speech length, and show better immunity to noise. In a specific continuous speech recognition testing system, favorable recognition result is obtained.