噪声下差分复合子带语音识别方法

本文根据子带特征反映语音信号局部特性和全带特征反映语音信号整体特性的事实，提出了 一种差分复合子带语音识别新方法。先用频谱差分减少噪声的干扰，再将多子带特征识别概率与全带特征识别概率相结合进行综合判决，以得到最终识别结果。将新方法应用于TIMIT数据包0－9十个英文数字和E－Set在NoiseX92的白噪声和F16战机噪声下的识别实验。实验结果表明新方法比传统方法识别性能有很大提高。

A differential hybrid subband method for noisy speech recognition

In this paper we present a differential hybrid subband method for noisy speech recognition based on facts that subband feature describes local speech characteristics and full band feature describes whole speech characteristics. We employ spectral difference method to reduce the disturbance of noise, and combine subband with full band at probability layer. The new method is evaluated by tasks on E-SET database and TIMIT digits database including ten English digits from 0 to 9. White noise and F16 noise of NoiseX92 database are added to original speech to simulate noisy speech at different SNRs. The experimental results show that the performance of speech recognition system can be improved greatly by the new method under noisy environment.