README.md

---
license: other
license_name: nscl-a2sb-and-polyform-nc
license_link: https://raw.githubusercontent.com/NVIDIA/diffusion-audio-restoration/refs/heads/main/LICENSE
tags:
  - audio
  - audio-restoration
  - schrodinger-bridge
  - diffusion
  - festival-audio
  - non-commercial
library_name: pytorch
pipeline_tag: audio-to-audio
---

# Soundboard

Schrödinger Bridge denoiser fine-tuned for musical recording audio restoration —
recovers a soundboard-style mix from heavily-corrupted audience recordings
(room reverb + audience-mic blend + lossy codec artifacts).

Fine-tuned from NVIDIA's
[A2SB](https://huggingface.co/nvidia/audio_to_audio_schrodinger_bridge)
(`twosplit_0.5_1.0` split) on a synthetic-corruption training pipeline driven
by **profile-based augmentation** — corruption parameters are calibrated
from real (clean, festival-recording) pairs and sampled at training time
from the recovered distribution. See [Locutius](https://github.com/protodotdesign/locutius)
for the full corruption chain, profiling, and training scaffold.

## Quick facts

| | |
|---|---|
| Architecture | AttnUNetF (565.5M params) |
| Audio format | 44.1 kHz, 2-channel, 32-bit float |
| Segment length | 130560 samples (2.96 s) |
| STFT | n_fft=2048, hop=512, window=hann |
| Representation | 3-channel `[mag^0.25, cos(phase), sin(phase)]` |
| Trained at step | 50,000 |
| Base checkpoint | NVIDIA A2SB `twosplit_0.5_1.0` |
| Checkpoint size | 2.1 GB |
| Diffusion | Schrödinger Bridge, β_max=1.0 |

## Usage

Load with the [Locutius](https://github.com/protodotdesign/locutius)
training package:

```python
import torch
from huggingface_hub import hf_hub_download
from locutius_train.config import TrainConfig
from locutius_train.network import AttnUNetF, SinusoidalTemporalEmbedding
from locutius_train.diffusion import Diffusion
from locutius_train.representation import WaveformToInput, InputToWaveform
from locutius_train.restore import restore_spectrogram

ckpt_path = hf_hub_download(repo_id="protodotdesign/Soundboard", filename="model.pt")
sd = torch.load(ckpt_path, map_location="cuda", weights_only=False)

cfg = TrainConfig()
model = AttnUNetF(
    n_updown_levels=cfg.model.n_updown_levels,
    in_channels=cfg.model.in_channels,
    hidden_channels=list(cfg.model.hidden_channels),
    out_channels=cfg.model.out_channels,
    emb_channels=cfg.diffusion.n_timestep_channels,
    band_embedding_dim=cfg.model.band_embedding_dim,
    n_attn_heads=cfg.model.n_attn_heads,
    attention_levels=list(cfg.model.attention_levels),
    use_attn_input_norm=cfg.model.use_attn_input_norm,
    num_res_blocks=cfg.model.num_res_blocks,
).to("cuda").eval()
model.load_state_dict(sd["model"])
```

See `restore.py` in the Locutius repo for a complete CLI that takes a
clean source, applies the calibrated festival-corruption profile, and
runs the reverse Schrödinger Bridge to produce a restored output.

## Calibrated corruption profile

This model was trained against a single calibrated profile recovered
from a real (studio FLAC, festival M4A) pair via per-kick local
Wiener deconvolution. The profile is bundled in `profile.json`:

```json
{
"name": "edc_festival",
"ir_path": "../impulses/EchoThief/Brutalism/San Diego Supercomputer Center Outdoor Patio California.wav",
"delay_ms_range": [
15.0,
25.0
],
"studio_gain_range": [
0.6,
0.7
],
"room_gain_range": [
0.55,
0.65
]
}
```

Each training-step corruption draws fresh values from these ranges,
so the model has been exposed to ~50,000 distinct delay/blend
combinations within the same venue character.

## Training data

Trained on a focused subset of electronic music FLACs. **No festival
recordings or other licensed audio were stored or distributed** —
only the studio source material was used; festival-corrupted versions
were synthesized on-the-fly from the calibrated profile during each
training step.

## Limitations

- **Single profile**: trained against one calibrated venue (`edc_festival`).
  Performance on festival recordings from very different venues / mix
  chains will degrade.
- **Electronic music bias**: training set was EDM-heavy. Restoration
  quality on rock, classical, or vocal-led material may be uneven.
- **No crowd-noise model**: the calibrated profile didn't include
  additive crowd-noise (no real crowd recordings were available
  during calibration). Recordings with heavy crowd vocals may have
  residual artifacts.
- **Non-commercial use only** — see the license below.

## License

Dual non-commercial license:

- [NVIDIA Source Code License for A2SB](LICENSE.NSCL-A2SB) (the upstream
  license inherited from the A2SB base checkpoint)
- [PolyForm Noncommercial 1.0.0](LICENSE.PolyForm-NC) (additional terms
  on top, source-availability + patent retaliation)

You must comply with **both** licenses. Use is restricted to research
and evaluation only — no commercial use is permitted. See
[LICENSING.md](https://github.com/protodotdesign/locutius/blob/main/LICENSING.md)
for the full plain-English breakdown.

## Citation

If you use this model in research, please cite the upstream A2SB paper
and reference this fine-tune:

```bibtex
@misc{soundboard,
  title={Soundboard: festival audio restoration via profile-calibrated Schrödinger Bridge fine-tuning},
  author={Locutius},
  year={2026},
  howpublished={\url{https://huggingface.co/protodotdesign/Soundboard}},
}
```