Joint estimation of vocal tract filter and glottal source waveform via convex optimization

For singing synthesis applications, it is essential to generate natural voiced sound, since singing is nearly 95% voiced. In this paper, the linearly separable source-filter model is assumed for voiced sound. A convex optimization method is then applied to estimate the vocal tract filter and glottal source waveform jointly. The merits of using convex optimization methods are computational efficiency and global optimality. The final results show that the joint estimation is effective in conjunction with a smoothing constraint.