# Fish Speech — PT-BR Fine-tuning (Paraíba Accent)

LoRA fine-tuning of **Fish Speech 1.5** ([fishaudio/fish-speech-1.5](https://huggingface.co/fishaudio/fish-speech-1.5), LLaMA component) on the PT-BR corpus `PB_PARAIBA` (May 2026). Checkpoints and configs at the root; `graficos/` with curves and metrics; `inferencias/` with evaluation audio.

## Checkpoints

| Path | Description |
|---|---|
| `final_merged_llama/` | Final model with LoRA merged (`model.pth` + tokenizer) — **use this for inference** |
| `merged_step_000010488/` | Merge of the checkpoint at step 10,488 |
| `lora_step_000010488.ckpt` | Raw LoRA checkpoint (Lightning) |

## Training

| Parameter | Value |
|---|---|
| Base model | fish-speech-1.5 |
| Steps | 10,488 (1 synthetic epoch) |
| Training time | ~44.7 h |
| Config | `training_config.yaml` |

## Files

- `graficos/` — train/eval loss curves and timing (PNG + CSV)
- `inferencias/` — `baseline_phrase_*.wav` × `step_000010488_phrase_*.wav` (baseline vs. fine-tuned), `reference_prompt.wav` + transcript, and `synthesis_phrases_ptbr.txt`
- `machine_profile.json`, `timing_*.json`, `dataset_text_stats.json` — environment, timing, and dataset stats

## Usage

Use `final_merged_llama/` as the LLaMA weights directory in the official [Fish Speech](https://github.com/fishaudio/fish-speech) pipeline.

**License:** derived from Fish Speech, subject to CC-BY-NC-SA-4.0.

## Citation

These checkpoints are LoRA fine-tunes of [fishaudio/fish-speech-1.5](https://huggingface.co/fishaudio/fish-speech-1.5). 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.