A Non-Fixed-Length Sequences Clustering Approach for Speech Corpus Reduction

In order to obtain better match among real variable length pitch sequences, this paper presents a non-fixed-length sequences clustering approach by using dynamic programming. Experimental results show that it can obtain better clustering result comparing with traditional methods. Combining with the speech coding technique, the size of the speech corpus can be significantly reduced and the obtained small-size multi-sample tonal mono-syllable corpus can satisfy the demands of clarity and naturalness for embedded text-to-speech systems.