SynPaFlex-Corpus: An Expressive French Audiobooks Corpus dedicated to expressive speech synthesis.

SynPaFlex-Corpus: An Expressive French Audiobooks Corpus Dedicated to Expressive Speech Synthesis Aghilas Sini1, Damien Lolive1, Gaëlle Vidal1, Marie Tahon3 and Élisabeth Delais-Roussarie2 1IRISA/Université de Rennes 1, Lannion, France; 2UMR 6310 - Laboratoire de Linguistique de Nantes, Nantes, France; 3 LIUM / Le Mans Université, Le Mans, France. elisabeth.delais-roussarie@univ-nantes.fr, marie.tahon@univ-lemans.fr {aghilas.sini, damien.lolive, gaelle.vidal}@irisa.fr Abstract This paper presents an expressive French audiobooks corpus containing eighty seven hours of good audio quality speech, recorded by a single amateur speaker reading audiobooks of different literary genres. This corpus departs from existing corpora collected from audiobooks since they usually provide a few hours of mono-genre and multi-speaker speech. The motivation for setting up such a corpus is to explore expressiveness from different perspectives, such as discourse styles, prosody, and pronunciation, and using different levels of analysis (syllable, prosodic and lexical words, prosodic and syntactic phrases, utterance or paragraph). This will allow developing models to better control expressiveness in speech synthesis, and to adapt pronunciation and prosody to specific discourse settings (changes in discourse perspectives, indirect vs. direct styles, etc.). To this end, the corpus has been annotated automatically and provides information as phone labels, phone boundaries, syllables, words or morpho-syntactic tagging. Moreover, a significant part of the corpus has also been annotated manually to encode direct/indirect speech information and emotional content. The corpus is already usable for studies on prosody and TTS purposes and is available to the community. Keywords: Speech corpus, audiobooks, emotion, expressiveness, 1. Introduction To build an expressive Text-To-Speech (TTS) system able to read books of different literary genres, using various discourse modes and speaking styles, a corpus that cov- ers all these specificities is required. Usually, corpora built for TTS purposes are less than ten hours long and mono- speaker. In addition, the content is carefully controlled to maximize the homogeneity of the synthetic speech. Long and coherent speech data is very interesting as it gives the possibility of studying voice expressiveness under dif- ferent situations. Audiobooks are a good example of such data and are valuable for prosody modeling, especially in the field of storytelling. For instance, (Montaño et al., 2013; Montaño Aparicio, 2016) show that expressive categories might exist in the storytelling speaking style. Some works have also been done to detect the speaking style in audio- books (Székely et al., 2012b) and to evaluate the usabil- ity of audiobooks data for the generation of conversational speech (Székely et al., 2012a). In the last decade, many corpora were built from audio- books. For instance, (Panayotov et al., 2015) presents a multi-speaker English corpus built for speech-text align- ment purposes containing thousands of hours of speech. In (Stan et al., 2013), a multilingual corpus that contains ap- proximately sixty hours of speech data from audiobooks in 14 languages is introduced. Since this corpus contains an average of four hours per language and only one book per language, it prevents studies on speaking styles partic- ularly for TTS synthesis. The same analysis is true for the GV-Lex corpus (Doukhan et al., …

Content truncated to 500 words