F5-TTS — PT-BR Fine-tuning (Paraíba Accent)

Fine-tuning of F5-TTS on a Brazilian Portuguese voice corpus (~41k segments), in three training runs. The fine-tuned runs started from the PT-BR checkpoint by firstpixel (firstpixel/F5-TTS-pt-br, firstpixelptbr/model_last.pt). Each training folder contains the checkpoint and configs at its root, graficos/ with curves and metrics, and inferencias/ with evaluation audio.

Training runs

Folder	Description	Checkpoints
`all/`	Full dataset, fine-tuned from firstpixel PT-BR	`model_last.pt` (step 443410)
`filtered/`	Quality-filtered dataset, fine-tuned from firstpixel PT-BR	`model_last.pt` (step 105852)
`filtered_zero_experiment/`	Trained from scratch (no pretrained weights), filtered dataset	`model_last.pt`, `model_2646300.pt`

Hyperparameters (filtered)

Parameter	Value
Starting checkpoint	firstpixel PT-BR (F5TTS_Base arch, finetune)
Learning rate	1e-05
Batch size	1000 frames/GPU
Epochs	6
Training samples	41,014
Tokenizer	custom (own vocab)

Full configuration in each run's setting.json and run_config.json.

Files (per training run)

model_*.pt — checkpoint
setting.json, run_config.json, statistics_summary.json — training configuration and statistics
graficos/ — loss curves (PNG) and per-step/per-epoch metrics (CSV)
inferencias/ — sample_step_*_{ref,gen}.wav (pairs generated during training), *frase*.wav/*line*.wav (per-checkpoint evaluation), and inferencias_*.csv

Usage

Load with the F5-TTS inference pipeline, using the custom vocab referenced in setting.json.

Citation

These checkpoints are fine-tunes of firstpixel/F5-TTS-pt-br (except the from-scratch run). 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.