Constructing Of The Embedded Mandarin Speech Corpus And Unit Searching

Corpus-Based speech synthesis approach has been applied in many state-of-art speech synthesis systems because it can produce synthetic voice of high naturalness. But high-quality speech synthesis system has high computing complexity and large resource consuming, so it is limited in the application on server and desk platform. In recent years along with the embedded technology's unceasing development, the application of embedded speech synthesis technology has become the inevitable tendency. But because the storage capacity and processing power of embedded system is limited, thus limiting the application of the speech synthesis technology in embedded systems. Therefore, how to resolve the conflict between the effect of synthesis and the consumption of resources, realize the applications of speech synthesis technology in the field of embedded system has become an urgent problem.Aiming at the small capacity and the finite capability of compute of embedded equipment, a transformative CART and improved K-center cluster based reducing algorithm, which can select the most representative units from the initial speech corpus to reduce the footprint of the units inventory, and an unit selection algorithm with low complexity are described in this paper.First, we take the accent syllable as the unit, adopt transformative CART to carry on pre-classification, and carry on the statistical analysis to the large corpus about the type of syllable, the total number of syllable and so on. And then, we respectively take the accent syllable and the initial/final as the unit, with frequency and duration, use improved K-center cluster algorithm to carry on cluster reducing at different proportion to get a tailored database which covers the most prosody situations of primary corpus. Finally, we use tailored database and improved selecting algorithm to build the speech synthesis system. We validate the algorithm of cluster reducing and unit selecting by carrying out subjective listening and objective testing. The result shows that the tailored database has high intelligibility and naturalness.