Adaptive Beamformer Based on Average Vowel / Consonant Spectrum with Phoneme Identification

For tele-conference systems or voice-controlled systems, the highquality sound capture of distant-talking speech is very important. However, background noise and room reverberations seriously degrade the sound capture quality in real acoustical environments. A microphone array is an ideal candidate for capturing distanttalking speech. With a microphone array, the desired speech signals can be acquired selectively by steering the directivity. Accordingly, a super-high directivity is necessary to reduce noise signals. To form directivity, delay-and-sum beamformer [1] and adaptive beamformers [2] [3] have been proposed as the conventional beamformers. A delay-and-sum beamformer forms the super-high directivity to the desired signal, and an adaptive beamformer forms null directivity to the noise signal. However, delay-and-sum beamformers have two serious drawbacks: the performance is not good enough to capture the desired signal without a sufficient number of transducers, and performance degrades in highly-reverberant rooms. On the other hand, adaptive beamformers can form null directivity with a small number of transducers. Furthermore, they can form sharper directivity than delay-and-sum beamformer. Consequently, adaptive beamformers are often used for the front-end processing of ASR (Automatic Speech Recognition). AMNOR (Adaptive Microphone-array for NOise Reduction) [3] is an adaptive beamformer proposed by Kaneda et al. in 1986. It promises a high quality sound-capture performance even in real acoustic environments. S-AMNOR [4] has also proposed in ICSLP2002. The S-AMNOR is the modified AMNOR based on a long time speech spectrum for capturing distant-talking speech with high quality. However, the S-AMNOR is not so suitable technique for recognizing the distant-talking speech, because speech has different characteristics as vowels and consonants. Therefore in this paper, we attempt to improve the speech recognition performance of the S-AMNOR with two adaptive filters based on vowel / consonant spectrum with phoneme identification.