Synthesising attitudes with global rhythmic and intonation contours

Abstract

We present here a trainable generative model of French prosody. We focus on the sentence level and design SNNs able to generate both rhythmic and intonation contours for diverse attitudes. First results of a perceptual test show that listeners are able to retrieve the right definition of attitudes by listening to synthetic PSOLA stimuli. 1. THEORETICAL FRAMEWORK In our theoretical framework prosody can be described as the superposition of independent multiparametric prosodic contours belonging to diverse linguistic levels [1]: sentence, clause, group, subgroup... These prototypical movements are progressively stored in a prosodic lexicon and dynamically used by the speaker to mark (segmentation...), enlight (salience) and enrich (attitudes...) the linguistic structuration of his discourse. In our approach, each syllable participates in the encoding of each linguistic level and higher levels can use whatever melodic or rhythmic variations to express linguistic representations. This theoretical framework contrasts with most popular models described in the literature: – Tonal approaches such as promoted by prosodic phonology where intonation is described with local events such as tones and breaks, the function of which are described [9] by higher phonological constructs such as the intonational, phonological phrase or word. – Superpositional models only based on physical or geometric [7] parameters such as cut-off frequencies or declination lines. – Data fitting or purely lexicon-based approaches, where synthesis is reduced to adequate and accurate labelling [4]. The model proposed here makes strong assumptions on the way linguistic and paralinguistic attributes are encoded in prosody. The main challenge of our work is to demonstrate that parameters of this model may be learned in order to adequately and accurately predict a multiparametric prosodic continuum. Input layer 3 F0 values (1/4tones) per IPCG IPCG Ratio 1 0 Number of syllables