acidente/fala_pb_checkpoints

Brazilian Portuguese TTS Checkpoints — Paraíba Accent

Fine-tuning checkpoints of five Text-to-Speech frameworks for Brazilian Portuguese (Paraíba accent), produced for the master's dissertation Incorporando Regionalismos e Sotaques em Modelos de Síntese de Fala (Thomaz Diniz, UFCG, 2026). All experiments share the same voice corpus (FALA_PB) and include checkpoints, evaluation audio, loss curves, and training reports.

Every checkpoint in this repository is a fine-tune of its respective base model (listed below and in the References section) — except F5TTS/filtered_zero_experiment/, which was trained from scratch as a control experiment.

Models

Folder	Base / starting checkpoint	Method	Size	Base license
F5TTS/	F5-TTS pt-br by firstpixel	Full fine-tuning (+ one from-scratch run)	~21 GB	CC-BY-NC-4.0
XTTS/	XTTS v2 (Coqui)	GPT fine-tuning	~37 GB	Coqui Public Model License
QwenTTS/	Qwen3-TTS-12Hz-1.7B-Base	Full fine-tuning	~13 GB	Apache-2.0
FishSpeech/	Fish Speech 1.5	LoRA + merge	~2.5 GB	CC-BY-NC-SA-4.0
OrpheusTTS/	Orpheus 3B (es_it) + ArtooDtoo PT-BR adapter	LoRA (Unsloth)	~800 MB	Llama 3 / Apache-2.0

Repository layout

Every model folder (or training folder, for F5TTS) follows the same pattern:

<Model>/
├── checkpoints          # .pt/.pth/.ckpt files or framework-format folders
├── configs & reports    # setting.json, REPORT.md, etc.
├── graficos/            # loss curves and metrics (PNG + CSV)
└── inferencias/         # evaluation audio (.wav) + phrases/manifests

• F5TTS/ has one extra level for its 3 training runs: all/, filtered/, filtered_zero_experiment/
• Folder-format checkpoints: QwenTTS/checkpoint-step-*/, FishSpeech/final_merged_llama/, OrpheusTTS/final_adapter/

Training data

Proprietary Brazilian Portuguese voice corpus with Paraíba accent (~41k segments, ~4 s average per sample). The dataset itself is not included in this repository — only the checkpoints, metrics, and generated evaluation audio.

Usage

Each model's README states the framework and version used for training. In short: load the checkpoint with the official inference pipeline of the corresponding framework (F5-TTS, Coqui TTS, Qwen3-TTS, Fish Speech, or Orpheus/Unsloth).

References — base models and starting checkpoints

The experiments started from the following models/checkpoints:

• F5-TTS: fine-tuned from the Brazilian Portuguese checkpoint by firstpixel — firstpixel/F5-TTS-pt-br (firstpixelptbr/model_last.pt), built on F5-TTS (SWivid). The filtered_zero_experiment run was trained from scratch (no pretrained weights).
• XTTS v2: fine-tuned from coqui/XTTS-v2 (Coqui AI).
• Qwen3-TTS: fine-tuned from Qwen/Qwen3-TTS-12Hz-1.7B-Base (Alibaba/Qwen).
• Fish Speech: LoRA fine-tuned from fishaudio/fish-speech-1.5 (Fish Audio).
• Orpheus TTS: LoRA fine-tuned with Unsloth from canopylabs/3b-es_it-ft-research_release (Canopy Labs), continuing from the PT-BR adapter ArtooDtoo/Orpheus_PTBR_FT_Unsloth_Chk-1800.

Citation

If you use these checkpoints, please cite the master's dissertation:

ABNT:

DINIZ, Thomaz. Incorporando regionalismos e sotaques em modelos de síntese de fala. 2026. Dissertação (Mestrado em Ciência da Computação) — Universidade Federal de Campina Grande, Campina Grande, 2026.

BibTeX:

@mastersthesis{diniz2026regionalismos,
  author  = {Diniz, Thomaz},
  title   = {Incorporando Regionalismos e Sotaques em Modelos de S{\'i}ntese de Fala},
  school  = {Universidade Federal de Campina Grande},
  address = {Campina Grande, Brazil},
  year    = {2026},
  type    = {Master's dissertation}
}