WAVELET TRANSFORM EXTREMA CLUSTERING FOR MULTI-CHANNEL SPEECH DEREVERBERATION

This paper presents a method for enhancing multi-channel reverber- ant speech using event-based processing of wavelet transform coef- ficients. Clustering of the wavelet extrema across multiple chan- nels is employed to obtain a single multi-scale extrema represen- tation from which the enhanced signal is synthesized. Processing is done in the LPC residual domain, with the entire analysis being preceded by a multi-channel LPC inverse filter and followed b y the corresponding forward LPC filter. The algorithm is compared to traditional delay-and-sum beamforming with results presented for reverberant, noisy conditions. In many real-world situations, speech signals are degraded by the reverberant effects of the environment in which they are pro duced. Traditional speech processing algorithms are designed to operate primarily under conditions where the noise is additive and uncorre- lated with the desired speech signal. Their effectiveness i s dramat- ically reduced in the presence of the non-stationary, convo lutional noise inherent in these reverberant settings. A number of me th- ods have been developed specifically to address the dereverb eration problem. Beamforming methods are effective at attenuating long term echoes (1) which tend to be uncorrelated across channels, but do little to reduce short term effects. More sophisticated a pproaches attempt to identify the channel effects in some form and compen- sate for them. These include cepstral processing (2), match ed fil- tering (3), and adaptive sub-space filtering (4). However, t he en- vironmental degradations in even a simple enclosure are very so- phisticated and quickly time-varying, and the temporal averaging required by these processes prohibits adaptation at a rate s ufficient to track and undo these effects. This imposes a fundamental limit on the effectiveness of this class of dereverberation techn iques. In (5) we proposed an alternative to these methods by explic- itly incorporating a speech model into the beamforming process. Our work in (6) extended this idea with an event-based process- ing method to discriminate impulses in the received signals due to channel effects from those present in the desired speech. These approaches were shown to suppress the deleterious effects of both reverberations and additive noise without explicitly iden tifying the channel and to be adaptable on a frame by frame basis. However, this earlier work was restricted to a small class of speech si gnals and limited to individual analysis frames. Here we take these concepts a step further by employing the wavelet domain for a multi-resolution signal analysis and reconstruction. This allows the model- based, event-based processing paradigm to be applied for effective speech enhancement under general conditions.