From segmental synthesis to acoustic rules using time dependent modeling techniques

Intelligible Text-to-Speech may be achieved by concatenating spectrally encoded segments. However, its Iack of naturalness could be attributed to a difficult control of speech parameters. Acoustic rules are more adequate for this control. The aim of this work is to provide a methodology to move from a segmental to a rule-based approach. A number of interactive tools is proposed using powerful signaland data analysis techniques for modeling spectral evolution, inferring spectral targets, and generating adequate transitions between these targets. The choice of adequate spectral parameters is essential. A set of French speech segments ("polysons ") of a single speaker has been encoded using these tools. Spectral targets were constrained to belong to a finite set of vectors (allophonic targets). Coarticulation effects (vowel reduction, nasalisation ... ) can be accounted for by controlling the time duration of temporal evolution functions. Segment. concatenation problems are eliminated. Automatie procedures to select allophonic targets for new speakers and group temporal pattems into rules are the current issues. UNLIMITED VOCABULARY SPEECH SYNTHESIS Two main approaches have been proposed for unlimited vocabulary speech synthesis. The segmental approach (using diphones, demi-syllables, "polysons", ... ) offers an easy way to intelligible speech. But the segment inventory is speaker dependent and control of timing is a non trivial task. Its lack of naturalness can be attributed to uneasy analytic control of speech parameters. A rule-based approach is more flexible, gives more insight on the relevant features of speech, and may allow speaker modification. Control of prosody, style of speech, sentence rythm, is achieved quite naturally within a unified framework. Unfortunately, this approach requires human knowledge and manual intervention, for visual and auditory hand-tuning of the rules. The time-rlependent modeling techniques, described below, permit a structural description of spectral evolution in speech segments. From their results, rules can then be obtained by inferring spectral targets and extracting typical transitions patterns between these targets. MODELING TEMPORAL EVOLUTION Speech can be encoded using a sequence of p-dimensional "spectral" vectors y(t), corresponding to a time sampling of the vocal tract transfer function. A synthesizer could, either retrieve from memory an appropriate sequence, or generate this sequence by rules. The rules should predict the vector y(t) for every time instant t. In order to infer these rules, two analysis techniques have been experimented with: vectorial AR models and temporal decomposition.