New pitch detection system for concurrent speech on a parallel processing system

Abstract

Pitch detection system based on the harmonic wavelet transform (HWT) algorithm for concurrent speech [Y. Chisaki, 6C.15, Proc. ICA2001] has been proposed. The HWT algorithm has two major advantages. One is to detect a pitch from a frame signal, which is a short duration, such as 20 ms. The other one is robustness against noise; namely, all pitches can be detected from concurrent speech in a short delay, simultaneously. However, processing time for the modified correlation method, as a conventional method is 8.6 times faster than that for the HWT based system. Due to the development of computers, processing time can be reduced easily by using PCs. This paper proposes a new pitch detection system for concurrent speech with a new parallel algorithm. The system is based on the HWT method, and implemented on a parallel processing system. Simulations are performed with some node arrangements. As a result, a processing speed for a parallelized block of the system is 6.3 times faster as that for the original one without being accompanied by degradation of pitch detection accuracy. [Part of this work is supported by The Sagawa Foundation for Promotion of Frontier Science, the Cooperative Research Project Program of the RIEC, Tohoku Univ., and the Ono Acoustics Research Fund (2001).]