用于抗噪声语音识别的谐振强度特征

Harmonic intensity feature for robust speech recognition

基于传统的Mel倒谱系数(MFCC)系列特征的语音识别系统在噪声环境中的识别性能会急剧下降。为了进行噪声环境中的自动语音识别,提出了一种反映语音信号谐振程度的特征:谐振强度,并用之代替传统MFCC特征中的能量维(零维倒谱C0,或者帧能量E)。在展览馆噪声、人群噪声和汽车噪声等情况下的语音识别实验结果表明:基于这种新特征的语音识别系统比基于传统特征的语音识别系统有更高的平均识别率和更好的抗噪声能力。

Automatic speech recognition (ASR) in noisy environments is a challenging problem. The performance of traditional Mel-frequency cepstral coefficient (MFCC) feature based ASR systems is dramatically degraded by additive noise. The harmonic intensity (H) feature was used to develop a robust ASR to replace the zero-order cepstral coefficient (C_0) or frame energy (E) feature in the MFCCs. A C_0-based ASR system, an E-based ASR system, and an H-based ASR system were tested with noise corrupted speech. The results show that the H-based ASR system has higher recognition accuracy and better robustness than the other systems.