# XTTS v2 — PT-BR Fine-tuning (Paraíba Accent)

GPT fine-tuning of **XTTS v2** ([coqui/XTTS-v2](https://huggingface.co/coqui/XTTS-v2)) on a Brazilian Portuguese voice corpus (April 2026). Checkpoints and configs at the root; `graficos/` with curves and metrics; `inferencias/` with evaluation audio.

## Checkpoints

| File | Step |
|---|---|
| `checkpoint_8868.pth` | 8,868 (1 epoch) |
| `checkpoint_17736.pth` | 17,736 |
| `checkpoint_26604.pth` | 26,604 |
| `checkpoint_35472.pth` | 35,472 |
| `checkpoint_44340.pth` | 44,340 |
| `best_model_53205.pth` | 53,205 (best eval loss) |
| `best_model.pth` | copy of the best model |

`config.json` holds the full training configuration; `experiment_config.json` and `relatorio_experimento.md` document the experiment.

## Training environment

- GPU: NVIDIA GeForce RTX 4060 Ti 16 GB
- PyTorch 2.5.1 + CUDA 12.1, Python 3.10, Windows 10
- Details in `machine_report_thesis.md`

## Files

- `*.pth`, `config.json`, `experiment_config.json` — checkpoints and configs
- `relatorio_experimento.md`, `training_statistics.json` — training reports
- `graficos/` — loss and training-time curves (PNG + CSV/JSON)
- `inferencias/` — `xtts_model_*_frase*.wav` (5 phrases per checkpoint, incl. `xtts_model_original_*` for comparison against the base model), `infer_finetuned_*` × `ref_*` pairs, and manifests
- `metadata_train.csv`, `metadata_eval.csv` — dataset manifests
- `xtts_finetune_paraiba_experiment.py` — experiment script

## Usage

Load with [Coqui TTS](https://github.com/coqui-ai/TTS) (`Xtts.init_from_config` + `load_checkpoint`). The original XTTS v2 support files (dvae, vocab, etc.) are not included — download them from the official Coqui repository.

**License:** derived from XTTS v2, subject to the [Coqui Public Model License](https://coqui.ai/cpml) (non-commercial use).

## Citation

These checkpoints are fine-tunes of [coqui/XTTS-v2](https://huggingface.co/coqui/XTTS-v2). If you use them, please cite:

> 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.