Vowel segregation in background noise using the model of segregating two acoustic sources

Masashi Unoki ${}^{* \dagger}$
^* ATR Human Information Processing
Research Laboratories
2-2, Hikaridai, Seika-cho, Soraku-gun,
Kyoto 619-0288, JAPAN
unoki@hip.atr.co.jp
Masato Akagi ${}^{\dagger}$
${}^{\dagger}$ School of Information Science, JAIST
1-1 Asahidai, Tatsunokuchi, Nomi-gun,
Ishikawa, 923-1292, JAPAN
akagi@jaist.ac.jp

Abstract:

This paper proposes an improved sound segregation model based on auditory scene analysis in order to overcome three disadvantages in our previously proposed model. The improved model solves the problem of segregating two acoustic sources by using constraints related to the heuristic regularities proposed by Bregman. In the improvements, we (1) reconsider the estimation of unknown parameters using Kalman filtering, (2) incorporate a constraint of channel envelopes with periodicity of the fundamental frequency into the grouping block, and (3) consider a constraint of smoothness of instantaneous amplitudes on channels. Simulations are performed to segregate a real vowel from a noisy vowel and to compare the results of using all or only some constraints. The proposed model can improve our previous model and precisely segregate real speech even in waveforms using all of the constraints related to Bregman's four regularities.