【摘要】 根据不同尺度子带特征反映语音的不同细节特性 ,提出一种噪声下的多层子带 ( MLS)语音识别方法。将语音频谱分成多层多个子带 ,首先各子带分别单独进行识别 ,然后将各层各子带识别概率综合起来得到最终识别结果。将新方法应用于 TIMIT数据包 E-Set在 Noise X92白噪声和 F1 6噪声下的识别实验。实验结果表明 ,多层子带方法在噪声环境和无噪情况下识别性能都有很大提高。更多还原

【Abstract】 A new method of combined multi-layer subband (MLS) for noisy speech recognition is presented, based on the fact that different scale subbands describe different characteristics of speech. Speech signal is divided into different multi-layer subbands. Each subband is trained and recognized respectively, and probabilities of all subbands are combined to form the final result. The new method is evaluated by a task on TIMIT database of E-Set. White noise and F16 noise of NoiseX92 are added to original speech at different SNRs. Experimental results show that the performance of speech recognition system can be improved by the new method in both noisy and clean environments.更多还原