Dept. of Information Eng., Kogakuin Univ., Tokyo, 163-91, Japan
NTT Advanced Technol. Corp.
MIT, Cambridge, MA 02138
Source waveform recovery under reverberant conditions is quite important for machine diagnostics. This talk investigates cepstrum dereverberation without using inverse filters of reverberant transfer functions. Source waveforms are typically modeled here by a pulselike signal, pulse trains, and modulated pulse trains. The dereverberation experiments demonstrate that a pulselike waveform, a periodic pulse train, and a modulated periodic train are recovered from the reverberant minimum phase cepstrum without nonminimum phase all-pass components. The all-pass components do not include important signature elements for those source signals. However, aperiodic pulse trains convey their signatures also in their nonminimum phase components. The recovery for those waveforms requires their all-pass components. The authors notice that speech waveforms are typical examples whose signature elements are included also in their nonminimum phase components due to their pitch fluctuations. The authors will also illustrate the zero locations of the source signals to describe not only the power spectra but also the phase property of the source signals. [Work supported by Sound Technology Foundation, Japan.]