Speech feature analysis and spectrum conversion from children to young adults

In this paper, the short time spectral envelope differences are analyzed and compared between children and young adults speech. Based on the analysis, a feature matching alignment Gaussian mixture model (FMA-GMM) is proposed to achieve the voice conversion from children to young adults. The model is gender-dependent and feature parallel training. In FMA-GMM, the F0 track matching degree is computed between several children and young adult speakers. Then a child and a young adult who have the best matching degree are chose to making feature warping alignment. The test speech is produced by twelve young people who provide the recordings in childhood. Experimental results show that the proposed method can achieve better performance than GMM and piece-wise linear warping function.