encoder_supertonic3_decoder/README.md

# BlueCodec encoder trained against the frozen official Supertonic-3 vocoder (Sept 2026)

| File | Content |
|------|---------|
| `encoder.safetensors` | Encoder weights only (`encoder.*` keys), for `BlueCodec.from_pretrained(..., decoder="supertonic3")` |
| `ae_290000.pt` | Training checkpoint at step 290,000 **without the decoder** (encoder, MPD, MRD, optimizers, schedulers; `decoder_source = "supertonic3"`), resumable with `train_autoencoder.py --encoder_only --decoder supertonic3 --resume` |
| `LICENSE.OpenRAIL-M` | Copy of the license of the Supertonic-3 model this encoder was trained against |

**The decoder is not in this repository.** This encoder uses the official Supertonic-3 vocoder (`onnx/vocoder.onnx` from [Supertone/supertonic-3](https://huggingface.co/Supertone/supertonic-3), revision `3cadd1ee`, Supertone Inc., BigScience OpenRAIL-M). The `bluecodec` package downloads it from the official repo at load time:

```python
from bluecodec import BlueCodec
codec = BlueCodec.from_pretrained("notmax123/blue-codec", decoder="supertonic3")   # needs: pip install onnx
z = codec.encode(audio, edge_pad_chunks=2)   # edge-padded encoding, recommended
y = codec.decode(z)[..., :audio.shape[-1]]
```

Recipe: encoder initialised from the 1.5M-step BlueCodec encoder, decoder frozen, AdamW lr 8.5e-5 cosine to 1e-6 over 300k steps, 2x64 segments of 61,740 samples, full-band log-mel reconstruction (λ 45) + LSGAN + composite feature matching, 10k-step discriminator warm-up. Details, audit and the end-of-clip edge code (use `edge_pad_chunks=2`): https://github.com/maxmelichov/blue-codec

**License note.** This encoder contains none of Supertone's weights, but it was trained through their model, so it may be a "Derivative of the Model" under OpenRAIL-M §1(e). Its use is therefore subject to the use-based restrictions in Attachment A of `LICENSE.OpenRAIL-M` in addition to the MIT license of BlueCodec's own code and weights.