OmniVoice BR-PT v1.5 is the selected Brazilian Portuguese fine-tune of k2-fsa/OmniVoice. This release uses the v1.5 gold-subset refinement checkpoint selected from checkpoint-9000.
It is optimized for Brazilian Portuguese TTS experiments, voice-design, and voice-cloning workflows using OmniVoice.
Why This Checkpoint
We evaluated base OmniVoice, the first BR-PT run, and v1.5 refine checkpoints on a 20-prompt Brazilian Portuguese ASR intelligibility set. Lower is better.
Model
WER
CER
base OmniVoice
0.0456
0.0400
v1 checkpoint-19000
0.0456
0.0400
v1.5 checkpoint-5000
0.0400
0.0379
v1.5 checkpoint-8000
0.0463
0.0404
v1.5 checkpoint-9000
0.0400
0.0379
v1.5 checkpoint-10000
0.0471
0.0390
checkpoint-9000 was selected because it tied for the best ASR score while being later in the refinement run than checkpoint-5000.
Samples
Reference audio was provided for sample generation and is included for reproducibility:
Gold refine set: 12,014 train rows, 260 dev rows after filtering
Speakers in selected gold subset: 40
Language ID used for OmniVoice compatibility: pt
Locale metadata: pt-BR
Instruction used in refine data/samples: portuguese accent
Refine LR: 5e-6
Refine steps: 10,000
Selected checkpoint: 9,000
Final refine eval loss at step 10,000: 3.93625
Usage
omnivoice-infer --model edwixx/omnivoice-brpt-v15 --text "Bom dia. Esta é uma demonstração em português brasileiro." --language pt --instruct "portuguese accent" --output brpt_v15.wav
Voice cloning example:
omnivoice-infer --model edwixx/omnivoice-brpt-v15 --text "Hoje eu vou mostrar uma novidade rápida e fácil de acompanhar." --language pt --instruct "portuguese accent" --ref_audio reference.wav --output brpt_clone.wav
Limitations
This is an experimental OmniVoice fine-tune. The model uses OmniVoice's generic Portuguese language code (pt) internally; Brazilian Portuguese behavior comes from the fine-tuning data, the gold subset, and consistent prompting. Human listening is still required to judge accent, naturalness, and voice-cloning quality.
Do not use this model for impersonation, deception, fraud, harassment, or cloning voices without consent.
Files
This repo intentionally excludes optimizer and random-state files to keep the release cleaner. Reports are in reports/, and audio examples are in samples/.
