# BVFM speech/text weights

## Files

- `bvfm_speech_step299999_inference.pt` (947,317,708 bytes): deployment-only
  checkpoint extracted from the selected step-299,999 training checkpoint. It
  contains `inference_modules`, model config, normalization/tokenizer/speaker
  metadata, and no optimizer, scaler, RNG, training-module duplicate, or EMA
  training container.
- `merged_config.json`: effective checkpoint configuration and training
  provenance. Runtime data/model locations can be overridden by the GitHub
  entry points.
- `semantic_vae_1000k/`: released Semantic-VAE EMA decoder metadata and
  weights used to decode the 40 Hz, 64-D speech latent.

## SHA-256

```text
6c4f1974e0e29ce7c4d755f1c49875eda5a8e4a03663d4289f0cd97e52df3429  bvfm_speech_step299999_inference.pt
7c455aa8ab3f7d576b4834f8342558894aafaa61a371b84a9bfa4d10a100e516  semantic_vae_1000k/dac/ema_state_dict.pth
```

With the GitHub code:

```bash
export BVFM_WEIGHTS_ROOT=/path/to/model-repo
python scripts/infer_tts_one.py \
  --ckpt-dir /path/to/model-repo/speech \
  --checkpoint bvfm_speech_step299999_inference.pt \
  --text "A short synthesis example." \
  --ref-wav /path/to/reference.wav \
  --out-dir runs/demo
```

The large resumable `latest.pt`, `final.pt`, and periodic step snapshots are
deliberately excluded from the Hugging Face package.