F0 prediction from linear predictive cepstral coefficient
Xueqin Chen, Yibiao Yu, Heming Zhao · 2014
In this paper, we proposed a fundamental frequency prediction method which is used primarily in the voice conversion system. This paper establishes a Gaussian Mixture Model (GMM) to predict the fundamental frequency based on the Linear Predictive Cepstral Coefficient (LPCC). The model may be the speaker-dependent Gaussian mixture model or the speaker-independent universal background model that is decided by the number of speakers in the training data set. The gross pitch error and the tone correct rate of the predicted F0 are reviewed and analyzed separately. The experimental results show that there is a relatively stable mapping relationship between LPCCs and fundamental. The gross pitch error in the GMM and UBM is about 7.52% and 14.03%. Subjective tests certify that the tone could be understood well. This F0 prediction method could be utilized to predict pitch in whispered speech conversion system and voice conversion system.