New Analysis Method For Harmonic Plus Noise Model Based on Time-Domain Periodicity Score

摘要

Harmonic plus Noise Model (HMN) is the two-band speech model which assumes the speech signal to be composed of a harmonic part and a noise part. It is very important to determine the reasonable maximum voiced frequency which delimit harmonic and noise parts for high-quality synthetic speech. In this paper, a new anlaysis method for HMM, especially maximum voiced frequency, is proposed, whereby the time-domain periodicity score is calculated for each harmonic band and the band of maximum score is selected for maximum voiced frequency. Moreover, recurrent equation of periodicity score is developed for fast implementation. This mehtod has proven to be robust to the inaccuracy of pitch estimation and the firequency resolution problem in low-pitched speech.