DEREVERBERATION OF SPEECH SIGNALS BASED ON SUBBAND ENVELOPE ESTIMATION

The full band inverse fiter of an acoustic field usually can not be approximated by causal filter because of its non-minimum phase property. However, by dividing the full band signal into many sub-bands, most of the sub-bands, that do not have zeros extremely close to or outside of the unit circle, can be inverted by causal filter. In this paper, a method of recovering the reverberated speech signals by sub-band envelope estimation is proposed. A reference signal and its reverberated version are supposed to be known at the parameter estimation phase. Both signals are divided into many sub-bands. The dereverberation filter for each sub-band is calculated by the least mean square estimation of the reference sub-band signal from the reverberated one. Reverberated signals under the same condition can be recovered using the estimated dereverberation filters. Comparison of the dereverberation error and distribution of the room transfer function zeros showed that the sub-bands that have large dereverberation errors are those with zeros extremely close to or outside of the unit circle. The performance of the dereverberation is improved with the increase of the number of sub-bands. The reverberated speech is recovered in good quality when the reverberation time is less than 0.43 s, which is approximately the reverberation condition of the general small office rooms.