Robust speech recognition based on emphasis filtering on formant regions in mobile noise environment

摘要

In real environment, the performance of automatic speech recognizer (ASR) has been degraded by various noises. Thus, some speech enhancement techniques have been applied but some distortions have still existed in the speech region. Actually, MFCC-derived features which have been popularly used in most of ASRs are strongly related with the spectral envelope. Moreover, the formant in the spectral envelope is very important factor to recognize the speech. However, it is very difficult to detect exactly the formant frequency in various noisy environments. Therefore, we propose the efficient algorithm to detect the spectral regions presenting any formants based on the spectral ratio instead of the direct detection of formant frequency. Then we check whether each detected frequency range is valid by several confidence measures. If it is acceptable, the region is amplified by the proper emphasis filter. The evaluation in the mobile voice search task shows that the proposed method provides some improvement over baseline system.