Removing Reflected Waves Using Temporal and Spectral Subtraction Without a priori Knowledge

KENKO OHTA,; DI PERSIA, LEANDRO EZEQUIEL; MASUZO YANAGIDA,

Sapporo

Congreso; IEEE International Workshop on Nonlinear Signal and Image Processing (NSIP 2005); 2005

IEEE

Proposed method is a method of removing reflectedwaves from a mixed wave consisting of a source signal andreflected waves. The method is a kind of waveform subtractionreferring to auto-correlation functions(ACFs) of multi channelspeech signals. We assume that a reflected wave has twoparameters; path amplitude and delay time. The methodestimates these parameters based on ACFs of signals receivedby microphones. The delay time of a particular path isestimated as the time lag that gives the maximum differencebetween the ACF of the channel in concern and the averageACF of the other rest channels. The delayed wave is subtractedfrom the received wave using an estimated delay only forvocalic segments and spectral subtraction is applied tonon-speech segments. The rate of waveform subtraction, or thepath amplitude of the reflected wave, is estimated byminimizing the difference between the ACF of the signal inconcern and the average ACF of the rests at the time delayattributed to the reflection in concern. The proposed methodcan be realized without prior knowledge about roomcharacteristics or the target speech. Speech recognition ratefor the signals picked up with 3 microphones in a reverberantenvironment is improved about 8% employing the proposedmethod.