Advocating for text input in multi-speaker text-to-speech systems
Résumé
Nowadays text-to-speech synthesis (TTS) systems are most
commonly trained using phonetic input. This is mostly due to
the poor performance of the letter-to-sound (L2S) mapping (in
particular with languages with opaque orthography) performed
by end-to-end TTS: the empirical distribution of the words sampled in the sole training corpus cannot compete with pronunciation dictionaries. Taylor and Richmond [1] actually reported
letter-to-sound errors – implicitly performed by end-to-end systems from raw text input – close to 10%.
This paper nevertheless shows that speakers produce lawful phonological variations and that end-to-end TTS systems
trained to accept text input – once trained adequately – can capture these variations of pronunciation that are strong markers of sociolinguistic features. We illustrate such variations on liaisons and schwas in French and r-linking in British English.
We therefore advocate for restoring text input for TTS, so that the many aspects of style variations (produced by speakers as well as stylistic variations) encoded by suprasegmental features can also be reflected in actual variations of pronunciation.
Origine : Fichiers éditeurs autorisés sur une archive ouverte