Using variable-sized speech segments as targets for concatenative Speech-to-Speech synthesis

摘要

Concatenative speech synthesis is growing in popularity due to the high naturalness of its resulting voice quality, but it is still domain-specific and has not yet been tested with conversational speech. We propose a method of unit selection that will overcome some of the problems that have prevented this development. In particular, we address two problems; one is the need for an extremely large database of labelled speech, the other is the incorporation of paralinguistic information in the speech synthesis. In our proposed 'speech-to-speech' method, we use acoustic criteria to segment the database into variable-sized units, and then use an acoustic waveform as a target for the unit-selection search. In a final stage, prosodic criteria are applied to select the optimal sequence of units for the output waveform generation. In this paper, we describe the techniques for segmenting the large speech database and the acoustic criteria used for unit selection. We present results comparing two methods of speech database segmentation, and further results from accuracy based on phonetic labels and a perceptual test which confirm the intelligibility and naturalness and accuracy of dictation.