Fix Readme
Browse files
README.md
CHANGED
|
@@ -1,157 +1,157 @@
|
|
| 1 |
-
|
| 2 |
-
|
| 3 |
-
|
| 4 |
-
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
|
| 102 |
-
|
| 103 |
}
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
|
| 107 |
-
|
| 108 |
-
|
| 109 |
-
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: nscl-a2sb-and-polyform-nc
|
| 4 |
+
license_link: https://raw.githubusercontent.com/NVIDIA/diffusion-audio-restoration/refs/heads/main/LICENSE
|
| 5 |
+
tags:
|
| 6 |
+
- audio
|
| 7 |
+
- audio-restoration
|
| 8 |
+
- schrodinger-bridge
|
| 9 |
+
- diffusion
|
| 10 |
+
- festival-audio
|
| 11 |
+
- non-commercial
|
| 12 |
+
library_name: pytorch
|
| 13 |
+
pipeline_tag: audio-to-audio
|
| 14 |
+
---
|
| 15 |
+
|
| 16 |
+
# Soundboard
|
| 17 |
+
|
| 18 |
+
Schrödinger Bridge denoiser fine-tuned for musical recording audio restoration —
|
| 19 |
+
recovers a soundboard-style mix from heavily-corrupted audience recordings
|
| 20 |
+
(room reverb + audience-mic blend + lossy codec artifacts).
|
| 21 |
+
|
| 22 |
+
Fine-tuned from NVIDIA's
|
| 23 |
+
[A2SB](https://huggingface.co/nvidia/audio_to_audio_schrodinger_bridge)
|
| 24 |
+
(`twosplit_0.5_1.0` split) on a synthetic-corruption training pipeline driven
|
| 25 |
+
by **profile-based augmentation** — corruption parameters are calibrated
|
| 26 |
+
from real (clean, festival-recording) pairs and sampled at training time
|
| 27 |
+
from the recovered distribution. See [Locutius](https://github.com/protodotdesign/locutius)
|
| 28 |
+
for the full corruption chain, profiling, and training scaffold.
|
| 29 |
+
|
| 30 |
+
## Quick facts
|
| 31 |
+
|
| 32 |
+
| | |
|
| 33 |
+
|---|---|
|
| 34 |
+
| Architecture | AttnUNetF (565.5M params) |
|
| 35 |
+
| Audio format | 44.1 kHz, 2-channel, 32-bit float |
|
| 36 |
+
| Segment length | 130560 samples (2.96 s) |
|
| 37 |
+
| STFT | n_fft=2048, hop=512, window=hann |
|
| 38 |
+
| Representation | 3-channel `[mag^0.25, cos(phase), sin(phase)]` |
|
| 39 |
+
| Trained at step | 50,000 |
|
| 40 |
+
| Base checkpoint | NVIDIA A2SB `twosplit_0.5_1.0` |
|
| 41 |
+
| Checkpoint size | 2.1 GB |
|
| 42 |
+
| Diffusion | Schrödinger Bridge, β_max=1.0 |
|
| 43 |
+
|
| 44 |
+
## Usage
|
| 45 |
+
|
| 46 |
+
Load with the [Locutius](https://github.com/protodotdesign/locutius)
|
| 47 |
+
training package:
|
| 48 |
+
|
| 49 |
+
```python
|
| 50 |
+
import torch
|
| 51 |
+
from huggingface_hub import hf_hub_download
|
| 52 |
+
from locutius_train.config import TrainConfig
|
| 53 |
+
from locutius_train.network import AttnUNetF, SinusoidalTemporalEmbedding
|
| 54 |
+
from locutius_train.diffusion import Diffusion
|
| 55 |
+
from locutius_train.representation import WaveformToInput, InputToWaveform
|
| 56 |
+
from locutius_train.restore import restore_spectrogram
|
| 57 |
+
|
| 58 |
+
ckpt_path = hf_hub_download(repo_id="protodotdesign/Soundboard", filename="model.pt")
|
| 59 |
+
sd = torch.load(ckpt_path, map_location="cuda", weights_only=False)
|
| 60 |
+
|
| 61 |
+
cfg = TrainConfig()
|
| 62 |
+
model = AttnUNetF(
|
| 63 |
+
n_updown_levels=cfg.model.n_updown_levels,
|
| 64 |
+
in_channels=cfg.model.in_channels,
|
| 65 |
+
hidden_channels=list(cfg.model.hidden_channels),
|
| 66 |
+
out_channels=cfg.model.out_channels,
|
| 67 |
+
emb_channels=cfg.diffusion.n_timestep_channels,
|
| 68 |
+
band_embedding_dim=cfg.model.band_embedding_dim,
|
| 69 |
+
n_attn_heads=cfg.model.n_attn_heads,
|
| 70 |
+
attention_levels=list(cfg.model.attention_levels),
|
| 71 |
+
use_attn_input_norm=cfg.model.use_attn_input_norm,
|
| 72 |
+
num_res_blocks=cfg.model.num_res_blocks,
|
| 73 |
+
).to("cuda").eval()
|
| 74 |
+
model.load_state_dict(sd["model"])
|
| 75 |
+
```
|
| 76 |
+
|
| 77 |
+
See `restore.py` in the Locutius repo for a complete CLI that takes a
|
| 78 |
+
clean source, applies the calibrated festival-corruption profile, and
|
| 79 |
+
runs the reverse Schrödinger Bridge to produce a restored output.
|
| 80 |
+
|
| 81 |
+
## Calibrated corruption profile
|
| 82 |
+
|
| 83 |
+
This model was trained against a single calibrated profile recovered
|
| 84 |
+
from a real (studio FLAC, festival M4A) pair via per-kick local
|
| 85 |
+
Wiener deconvolution. The profile is bundled in `profile.json`:
|
| 86 |
+
|
| 87 |
+
```json
|
| 88 |
+
{
|
| 89 |
+
"name": "edc_festival",
|
| 90 |
+
"ir_path": "../impulses/EchoThief/Brutalism/San Diego Supercomputer Center Outdoor Patio California.wav",
|
| 91 |
+
"delay_ms_range": [
|
| 92 |
+
15.0,
|
| 93 |
+
25.0
|
| 94 |
+
],
|
| 95 |
+
"studio_gain_range": [
|
| 96 |
+
0.6,
|
| 97 |
+
0.7
|
| 98 |
+
],
|
| 99 |
+
"room_gain_range": [
|
| 100 |
+
0.55,
|
| 101 |
+
0.65
|
| 102 |
+
]
|
| 103 |
}
|
| 104 |
+
```
|
| 105 |
+
|
| 106 |
+
Each training-step corruption draws fresh values from these ranges,
|
| 107 |
+
so the model has been exposed to ~50,000 distinct delay/blend
|
| 108 |
+
combinations within the same venue character.
|
| 109 |
+
|
| 110 |
+
## Training data
|
| 111 |
+
|
| 112 |
+
Trained on a focused subset of electronic music FLACs. **No festival
|
| 113 |
+
recordings or other licensed audio were stored or distributed** —
|
| 114 |
+
only the studio source material was used; festival-corrupted versions
|
| 115 |
+
were synthesized on-the-fly from the calibrated profile during each
|
| 116 |
+
training step.
|
| 117 |
+
|
| 118 |
+
## Limitations
|
| 119 |
+
|
| 120 |
+
- **Single profile**: trained against one calibrated venue (`edc_festival`).
|
| 121 |
+
Performance on festival recordings from very different venues / mix
|
| 122 |
+
chains will degrade.
|
| 123 |
+
- **Electronic music bias**: training set was EDM-heavy. Restoration
|
| 124 |
+
quality on rock, classical, or vocal-led material may be uneven.
|
| 125 |
+
- **No crowd-noise model**: the calibrated profile didn't include
|
| 126 |
+
additive crowd-noise (no real crowd recordings were available
|
| 127 |
+
during calibration). Recordings with heavy crowd vocals may have
|
| 128 |
+
residual artifacts.
|
| 129 |
+
- **Non-commercial use only** — see the license below.
|
| 130 |
+
|
| 131 |
+
## License
|
| 132 |
+
|
| 133 |
+
Dual non-commercial license:
|
| 134 |
+
|
| 135 |
+
- [NVIDIA Source Code License for A2SB](LICENSE.NSCL-A2SB) (the upstream
|
| 136 |
+
license inherited from the A2SB base checkpoint)
|
| 137 |
+
- [PolyForm Noncommercial 1.0.0](LICENSE.PolyForm-NC) (additional terms
|
| 138 |
+
on top, source-availability + patent retaliation)
|
| 139 |
+
|
| 140 |
+
You must comply with **both** licenses. Use is restricted to research
|
| 141 |
+
and evaluation only — no commercial use is permitted. See
|
| 142 |
+
[LICENSING.md](https://github.com/protodotdesign/locutius/blob/main/LICENSING.md)
|
| 143 |
+
for the full plain-English breakdown.
|
| 144 |
+
|
| 145 |
+
## Citation
|
| 146 |
+
|
| 147 |
+
If you use this model in research, please cite the upstream A2SB paper
|
| 148 |
+
and reference this fine-tune:
|
| 149 |
+
|
| 150 |
+
```bibtex
|
| 151 |
+
@misc{soundboard,
|
| 152 |
+
title={Soundboard: festival audio restoration via profile-calibrated Schrödinger Bridge fine-tuning},
|
| 153 |
+
author={Locutius},
|
| 154 |
+
year={2026},
|
| 155 |
+
howpublished={\url{https://huggingface.co/protodotdesign/Soundboard}},
|
| 156 |
+
}
|
| 157 |
+
```
|