Incorporando regionalismos e sotaques em modelos abertos de síntese de fala.

Abstract

Open-weight speech synthesis models show limitations regarding the diversity of regional accents of Brazilian Portuguese. This dissertation investigates the incorporation of the Paraíba accent into three architectures (XTTS, FishSpeech, and Orpheus TTS), comparing three adaptation strategies: voice cloning, fine-tuning, and training from scratch. Exploratory experiments revealed practical limitations of voice cloning and training from scratch, and fine-tuning, identified as the most promising alternative, was systematically validated. To this end, the FALA_PB corpus was built, with about 111 hours of Paraíba-accented speech organized into short segments, each paired with its corresponding sentence-level transcription. The corpus aggregates data from pre-existing databases (CoLingPB, Braccent, and Sotaque Brasileiro) and from a dedicated collaborative collection; in particular, the CoLingPB interviews were adapted for the synthesis task, with the audios segmented into sentences, automatically cut with FFmpeg, and filtered by acoustic quality. The fine-tuned models were evaluated through automatic indicators of quality and intelligibility and through a subjective evaluation with 56 listeners. The results indicate that fine-tuning was effective: the Paraíba accent was perceived as more present in the fine-tuned models in about 88% to 90% of the responses, with no perceptible loss of quality, with XTTS achieving the best overall performance. As contributions of this dissertation, all developed code and data artifacts are provided, in addition to a reproducible evaluation protocol, broadening regional representation in open speech synthesis for Brazilian Portuguese and supporting future research aimed at the automatic synthesis of regionalisms and accents from other Brazilian regions.