Voice Accompaniment Separation Eval

---
name: voice-accompaniment-separation-eval
description: Evaluates a model's ability to separate vocal and accompaniment tracks from mixed music audio. It probes long-term dependency modeling and pattern repetition exploitation in audio source separation. Use when the user wants to benchmark on DSD100, MedleyDB, CCMixer, or asks about evaluating this task. Reports SDR.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2003.08954
  bibtex_key: liu2020voice
  confidence: high
---

# voice-accompaniment-separation-eval

> Voice and accompaniment separation in music using self-attention convolutional neural network — Liu et al. (2020) (arXiv:2003.08954, 2020)

## What this evaluates

Evaluates a model's ability to separate vocal and accompaniment tracks from mixed music audio. It probes long-term dependency modeling and pattern repetition exploitation in audio source separation.

## Datasets

- **DSD100** — total ?; splits: test (25)
- **MedleyDB** — total ?; splits: val (-1), test (-1)
- **CCMixer** — total ?; splits: val (-1), test (-1)

## Metrics

- `SDR` **(primary)** — range: dB
  - Track-wise normalised Signal-to-Distortion Ratio in decibels (dB). Average results across the test set are reported.
- `SIR` — range: dB
  - Track-wise normalised Signal-to-Interference Ratio in decibels (dB). Average results across the test set are reported.
- `SAR` — range: dB
  - Track-wise normalised Signal-to-Artifact Ratio in decibels (dB). Average results across the test set are reported.

## Input / output format

**Input**: Mixed music audio tracks downsampled to 16 kHz.

**Output**: Separated vocal and accompaniment audio tracks.

## Scoring recipe

```python
# For each track in the test set:
# 1. Compute track-wise normalised SDR, SIR, and SAR between predicted and ground truth sources.
# 2. Average the metric values across all tracks in the test set.
# 3. Report mean SDR, SIR, SAR for both voice and accompaniment.
```

## Common pitfalls

- The test set is a custom combination of 1/3 MedleyDB, 1/3 CCMixer, and 25 DSD100 test tracks, not the standard DSD100 test set.
- All audio must be downsampled to 16 kHz before evaluation.
- Baselines like MMDenseLSTM may yield different results if trained with different structures or recipes, so direct comparison requires identical training setups.

## Evidence (verbatim from paper)

> We calculate the track-wise (normalised) SDR, SIR, and SAR as metrics. The average results on the test set are reported in Table I.

## Citation

```bibtex
@misc{liu2020voice,
  title={Voice and accompaniment separation in music using self-attention convolutional neural network},
  author={Liu et al. (2020)},
  year={2020},
  note={arXiv:2003.08954}
}
```

- arXiv: 2003.08954