Add paper and GitHub links, and sample usage
#2
by
nielsr
HF Staff
- opened
README.md
CHANGED
|
@@ -6,8 +6,26 @@ task_categories:
|
|
| 6 |
|
| 7 |
# MSRBench: A Benchmarking Dataset for Music Source Restoration
|
| 8 |
|
| 9 |
-
|
| 10 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 11 |
|
| 12 |
## Download
|
| 13 |
`hf download yongyizang/MSRBench --repo-type dataset --local-dir {your output dir}`
|
|
@@ -47,22 +65,7 @@ Every *.flac file is 48 kHz stereo; each {stem_name} contains 250 song_ids, corr
|
|
| 47 |
|
| 48 |
### DT1-4: Analog and Environmental Distortions
|
| 49 |
#### Radio (DT1)
|
| 50 |
-
Simulates FM broadcasting using GNU Radio's stereo multiplex encoding/decoding blocks with standard broadcast parameters
|
| 51 |
-
|
| 52 |
-
**Signal Processing Parameters:**
|
| 53 |
-
- Audio sampling rate: 44.1 kHz
|
| 54 |
-
- Quadrature rate: 220.5 kHz
|
| 55 |
-
- Pre-emphasis/de-emphasis time constant: 75 μs (US standard)
|
| 56 |
-
- Maximum frequency deviation: 75 kHz
|
| 57 |
-
- Stereo pilot tone: 19 kHz
|
| 58 |
-
|
| 59 |
-
**Channel Model:**
|
| 60 |
-
- **Rayleigh fading:** 8-sinusoid sum-of-sinusoids model with 20 Hz maximum Doppler shift
|
| 61 |
-
- **Carrier frequency offset:** σ = 0.1, max 1 kHz variations
|
| 62 |
-
- **Multipath channel:** 3-tap configuration
|
| 63 |
-
- Delays: [0, 0.1, 0.5] ms
|
| 64 |
-
- Relative magnitudes: [1.0, 0.8, 0.3]
|
| 65 |
-
- **Noise:** Additive white Gaussian noise at approximately 26 dB SNR (noise amplitude = 0.05)
|
| 66 |
|
| 67 |
#### Cassette (DT2)
|
| 68 |
Models magnetic tape coloration and noise using the DAW Cassette plugin by Klevgrand.[^1]
|
|
@@ -71,32 +74,31 @@ Models magnetic tape coloration and noise using the DAW Cassette plugin by Klevg
|
|
| 71 |
Reproduces playback artifacts including crackle and mechanical noise using iZotope Vinyl,[^2] configured with the "1970" preset.
|
| 72 |
|
| 73 |
#### Live Sound (DT4)
|
| 74 |
-
Generated using
|
| 75 |
-
- Impulse responses from PyRoomAcoustics
|
| 76 |
-
- Bandpass filtering to approximate phone microphone characteristics
|
| 77 |
-
- Environmental noise from the WHAM! dataset 48 kHz version mixed at approximately 20 dB SNR
|
| 78 |
|
| 79 |
### DT5-8: Traditional Lossy Audio Codecs
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
#### DT7: 128 Kbps AAC
|
| 86 |
-
|
| 87 |
-
#### DT8: 128 Kbps MP3
|
| 88 |
|
| 89 |
### DT9-12: Neural Audio Codecs
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 98 |
|
| 99 |
---
|
| 100 |
|
| 101 |
[^1]: [Klevgrand DAW Cassette](https://klevgrand.com/products/dawcassette)
|
| 102 |
-
[^2]: [iZotope Vinyl](https://www.izotope.com/en/products/vinyl)
|
|
|
|
| 6 |
|
| 7 |
# MSRBench: A Benchmarking Dataset for Music Source Restoration
|
| 8 |
|
| 9 |
+
[**Project Page**](https://msrchallenge.com/) | [**Paper**](https://huggingface.co/papers/2601.04343) | [**Code**](https://github.com/yongyizang/MSRKit)
|
| 10 |
|
| 11 |
+
MSRBench is the validation dataset for the MSR Challenge 2025; it contains 250 professionally mixed audio clips and their corresponding ground-truth targets under both original mix and 12 degradation scenarios in three categories.
|
| 12 |
+
|
| 13 |
+
## Sample Usage
|
| 14 |
+
|
| 15 |
+
The official [MSRKit](https://github.com/yongyizang/MSRKit) repository provides scripts for inference and evaluation on this dataset.
|
| 16 |
+
|
| 17 |
+
### Inference
|
| 18 |
+
To run inference on audio files using a pretrained model:
|
| 19 |
+
```bash
|
| 20 |
+
python inference.py --config config.yaml --checkpoint path/to/your/checkpoint.ckpt --input_dir path/to/your/input/directory --output_dir path/to/your/output/directory
|
| 21 |
+
```
|
| 22 |
+
|
| 23 |
+
### Evaluation
|
| 24 |
+
To calculate metrics (Multi-Mel-SNR, Zimtohrli, FAD-CLAP):
|
| 25 |
+
```bash
|
| 26 |
+
python calculate_metrics.py {file list}
|
| 27 |
+
```
|
| 28 |
+
*Note: The evaluation script expects a file list with each line in the format of `{target path}|{output path}`.*
|
| 29 |
|
| 30 |
## Download
|
| 31 |
`hf download yongyizang/MSRBench --repo-type dataset --local-dir {your output dir}`
|
|
|
|
| 65 |
|
| 66 |
### DT1-4: Analog and Environmental Distortions
|
| 67 |
#### Radio (DT1)
|
| 68 |
+
Simulates FM broadcasting using GNU Radio's stereo multiplex encoding/decoding blocks with standard broadcast parameters.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 69 |
|
| 70 |
#### Cassette (DT2)
|
| 71 |
Models magnetic tape coloration and noise using the DAW Cassette plugin by Klevgrand.[^1]
|
|
|
|
| 74 |
Reproduces playback artifacts including crackle and mechanical noise using iZotope Vinyl,[^2] configured with the "1970" preset.
|
| 75 |
|
| 76 |
#### Live Sound (DT4)
|
| 77 |
+
Generated using impulse responses from PyRoomAcoustics, bandpass filtering, and environmental noise from the WHAM! dataset.
|
|
|
|
|
|
|
|
|
|
| 78 |
|
| 79 |
### DT5-8: Traditional Lossy Audio Codecs
|
| 80 |
+
- **DT5:** 64 Kbps AAC
|
| 81 |
+
- **DT6:** 64 Kbps MP3
|
| 82 |
+
- **DT7:** 128 Kbps AAC
|
| 83 |
+
- **DT8:** 128 Kbps MP3
|
|
|
|
|
|
|
|
|
|
|
|
|
| 84 |
|
| 85 |
### DT9-12: Neural Audio Codecs
|
| 86 |
+
- **DT9:** 22 kHz Descript Audio Codec (DAC)
|
| 87 |
+
- **DT10:** 44 kHz DAC
|
| 88 |
+
- **DT11:** 6 Kbps Encodec
|
| 89 |
+
- **DT12:** 3 Kbps Encodec
|
| 90 |
+
|
| 91 |
+
## Citation
|
| 92 |
+
```bibtex
|
| 93 |
+
@article{msrchallenge2025,
|
| 94 |
+
title={Summary of The Inaugural Music Source Restoration Challenge},
|
| 95 |
+
author={Zang, Yongyi and others},
|
| 96 |
+
journal={arXiv preprint arXiv:2601.04343},
|
| 97 |
+
year={2026}
|
| 98 |
+
}
|
| 99 |
+
```
|
| 100 |
|
| 101 |
---
|
| 102 |
|
| 103 |
[^1]: [Klevgrand DAW Cassette](https://klevgrand.com/products/dawcassette)
|
| 104 |
+
[^2]: [iZotope Vinyl](https://www.izotope.com/en/products/vinyl)
|