Speech parameter generation considering LSP ordering property for HMM-based speech synthesis

LSP has many advantages for speech representation, especially correlates well to spectrum formants as long as the LSP parameters are strictly ordered and bounded. This ordering property cannot be guaranteed during HMM-based speech synthesis when LSP is adopted as the spectrum feature, because diagonal covariance is utilized and correlation between LSP dimensions is ignored, with the result that unstable issue will be caused in synthesized speech. In this paper, we modify the parameter generation criterion to preserve ordering property of generated LSPs, by considering not only the likelihoods for HMM and GV maximized in conventional method but also a mis-orderings penalty. Experimental results show that the proposed method can alleviate the mis-orderings significantly and achieve high quality synthesizing performance when the penalty weight is selected appropriately.