A pitch correction system based on spectral modeling synthesis technique

Abstract

Pitch correction with preservation of voice timbre has been a challenging issue that hinders the development of music related applications. To address this problem, we propose a spectral modeling synthesis (SMS) based automatic pitch correction system for singing voice. When obtaining one candidate singing voice, we firstly conduct pitch detection with voiced/unvoiced decision, and extracts the pitch contour. Then it is compared with the reference pitches that is extracted from the singing voice of famous singer, such that the detuned parts can be determined. We perform SMS method to obtain the frequency components together with the spectral envelop of the candidate singing voice, and then continue adjusting the phase of every spectral peak by preserving the spectral envelop, until all the pitches are corrected. Experimental results show that the proposed system can provide good pitch correction effect.