Acoustic Modeling for Lhasa Tibetan Speech Synthesis Based on DAEM Algorithm

Abstract

This paper applies the deterministic annealing expectation maximum (DAEM) algorithm into HMM-based Lhasa Tibetan speech synthesis. In this way, we can synthesize Lhasa Tibetan speech with non-time labeled training speech corpus. EM algorithm has the problem of initialization dependence, which can cause the problem of local maximum values. The DAEM algorithm has been introduced to solve this problem. In the process embedded re-evaluation during the model training, this method can make the computer obtain the optimal parameters to determine the best time boundary of each speech synthesis unit. Objective and subjective evaluation show that the synthesized Lhasa Tibetan speech has similar quality with the synthesized speech with time labeled speech corpus. Especially, the MOS of the synthesized Tibetan speech by the DAEM algorithm is higher when the number of training sentences is 700. Therefore, proposed method can be used for training acoustic models of Lhasa Tibetan speech synthesis with non-time labeled training speech corpus.