IHPLEHENTATION OF SHORT-TIME llOMOHORPHIC DEREVERBERATION

For short rp-verberation times (~ 10 msec) t and over short intervals (~ 50 msec) , speech waveforms reverberated by small rooms can be approximated by the convolution of a function interpreted as the "room impulse response" with the uncorrupted speech. It is possible, therefore, to remove the effects of reverberation by applying a deconvolution process to 50-msec sections of reverberated waveform. Homomorphic deconvolutiont ,.,hichconsists of transforming the convolved signal components into a sum of corresponding signals, linear-filtering the sum, and reconstructing dereverberated speech from the result, is well-suited to this application. Computer simulations discussed in this paper show that such a process is effective for artificially-reverberated speech. Neu results include filtering techniques that are useful when the reverberation is not precisely known before processing. THESIS SUPERVISOR: Alan V. Oppenheim TITLE: Associate Professor of Electrical Engineering