Suitable Reverberation Criteria for Distant-talking Speech Recognition

Abstract

The recognition of distant-talking speech has rapidly improved in recent years, because many novel speech-recognition techniques have been proposed that are robust against noise and reverberance. The signal to noise ratio (SNR) is generally used as a common criterion in speech-recognition techniques that are robust against noise. SNR is an effective noise criterion for estimating the recognition of speech in noisy environments. As an algorithm based on the perceptual evaluation of speech quality (PESQ) (T. Yamada et al. (2006)) has also been proposed to achieve the same target, we can roughly estimate the recognition of speech in noisy environments. However, no common reverberation criteria have been proposed to attain robust reverberant-speech recognition. It has therefore been difficult to estimate the recognition of reverberant speech. The reverberation time, T60, (M. R. Schroeder (1965)) is currently generally used to recognize distant-talking speech as a reverberation criterion. It is unique and does not depend on the position of the source in a room. However, distant-talking speech recognition greatly depends on the location of the talker relative to that of the microphone and the distance between them. Therefore, T60 is unsuitable for measuring the recognition of distant-talking speech. We propose newly reverberation criteria for measuring the recognition of distant-talking speech to overcome this problem. We first investigate suitable reverberation criteria to enable distant-talking speech to be recognized. We calculated automatic speech recognition with early and late reflections based on the impulse response between a talker and the microphone. We then evaluated it based on ISO3382 acoustic parameters (ISO3382 (1997)). Based on above investigation, we finally propose novel reverberation criteria RSR-Dn (Reverberant Speech Recognition criteria with Dn) which utilise ISO3382 acoustic parameters for robustly estimating reverberant speech recognition performance.