Suppression of speech intelligibility loss through a modulation transfer function-based speech dereverberation method

Abstract

The concept of modulation transfer function (MTF) can successfully be applied to evaluate the quality of speech transmission in the room acoustics [Houtgast and Steeneken, J. Acoust. Soc. Am. 77, 1069–1077 (1985)]. We previously proposed a speech dereverberation method based on the MTF concept, which consisted of MTF-based power envelope inverse filtering and the carrier regeneration in the filterbank [Unoki et al., EuroSpeech2003 (2003)]. This paper evaluates how the proposed method can suppress the loss of speech intelligibility caused by reverberation, by comparing various methods. We have carried out massive simulations of dereverberation for reverberant speech signals to objectively evaluate these methods. We also subjectively evaluated the methods via listening tests. In these simulations, artificial reverberations were convolved with a clean speech signal in which the impulse responses in the room acoustics can be approximated from the exponential decay as a function of reverberation time with a white-noise carrier. The results of both evaluations show that, in addition to reducing the averaged log-spectrum distortion by about 1 dB, the proposed method reduces the loss of speech intelligibility by about 30%. [Work supported by a Grant-in-Aid for Science Research from the Japanese Ministry of Education (No. 18680017).]