Time-domain concatenative text-to-speech synthesis.

Abstract

A concatenation framework for time-domain concatenative speech synthesis (TDCSS) is presented and evaluated. In this framework, speech segments are extracted from CV, VC, CVC and CC waveforms, and abutted. Speech rhythm is controlled via a single duration parameter, which specifies the initial portion of each stored waveform to be output. An appropriate choice of segmental durations reduces spectral discontinuity problems at points of concatenation, thus reducing reliance upon smoothing procedures. For text-to-speech considerations, a segmental timing system is described, which predicts segmental durations at the word level, using a timing database and a pattern matching look-up algorithm. The timing database contains segmented words with associated duration values, and is specific to an actual inventory of concatenative units. Segmental duration prediction accuracy improves as the timing database size increases. The problem of incomplete timing data has been addressed by using `default duration' entries in the database, which are created by re-categorising existing timing data according to articulation manner. If segmental duration data are incomplete, a default duration procedure automatically categorises the missing speech segments according to segment class. The look-up algorithm then searches the timing database for duration data corresponding to these re-categorised segments. The timing database is constructed using an iterative synthesis/adjustment technique, in which a `judge' listens to synthetic speech and adjusts segmental durations to improve naturalness. This manual technique for constructing the timing database has been evaluated. Since the timing data is linked to an expert judge's perception, an investigation examined whether the expert judge's perception of speech naturalness is representative of people in general. Listening experiments revealed marked similarities between an expert judge's perception of naturalness and that of the experimental subjects. It was also found that the expert judge's perception remains stable over time. A synthesis/adjustment experiment found a positive linear correlation between segmental durations chosen by an experienced expert judge and duration values chosen by subjects acting as expert judges. A listening test confirmed that between 70% and 100% intelligibility can be achieved with words synthesised using TDCSS. In a further test, a TDCSS synthesiser was compared with five well-known text-to-speech synthesisers, and was ranked fifth most natural out of six. An alternative concatenation framework (TDCSS2) was also evaluated, in which duration parameters specify both the start point and the end point of the speech to be extracted from a stored waveform and concatenated. In a similar listening experiment, TDCSS2 stimuli were compared with five well-known text-tospeech synthesisers, and were ranked fifth most natural out of six.