Adaptive beamformer with vowel/consonant identification based on average vowel/consonant spectrum for nosiy speech recognition

摘要

A microphone-array is an ideal candidate for capturing distant-talking speech. The AMNOR (Adaptive Microphone-array for NOise Reduction) is an adaptive beamformer proposed by Kaneda, et. al. In addition, as the beamformer for speech capture, S-AMNOR, the AMNOR with a long time speech spectrum was also proposed by Okada, et. al. In this paper, we propose the new AMNOR with adaptive filters for vowels/consonants, in order to improve the signal capturing performance. In addition, we automatically identify the vowels and consonants from output signal of the AMNOR by using a GMM (Gaussian Mixture Model). The performance of the proposed system is evaluated through experiments with a microphone array in a real room. As a result of evaluation experiments, the ASR (Automatic Speech Recognition) performance of proposed method achieved 65% in SNR = 10dB (speech comes from 90 deg. and noise comes from 50 deg.) environment, although the ASR performance of the conventional method is 60%.