由局部能量确定可靠性加权的抗噪声语音识别方法

Local Energy Based Reliability Weight for Robust ASR

噪声情况下的语音识别是个挑战性的问题。目前的处理方法普遍需要估计噪声或者信噪比,从而其性能依赖于噪声估计的好坏。本文提出了一种基于语音信号局部能量的可靠性加权方法,该方法着眼于语音本身的结构,避免了对噪声的估计。另外,带噪语音识别的实验结果证明该方法能很好的提高识别系统的抗噪声性能。

Automatic Speech Recognition (ASR) under noisy environment is still a challenging problem. Most resent approaches to the problem employ noise estimate or Signal-Noise Ratio (SNR) estimate, and hence their performances depend on the estimation of noise. In this paper, by introducing the local energy based reliability weight into the likelihood calculation, a novel approach is proposed, which focuses on the structure of speech rather than noise, and avoids the noise estimation. Moreover, the experimental results show that the noise-robustness of recognition systems can be manifestly improved with the proposed approach.