Wave Unet Musdb Eval

---
name: wave-unet-musdb-eval
description: Evaluates end-to-end time-domain audio source separation models on singing voice and multi-instrument separation tasks. Probes the model's ability to isolate specific audio sources from mixed recordings using raw waveform inputs. Use when the user wants to benchmark on MUSDB, CCMixter, or asks about evaluating this task. Reports MSE.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 1806.03185
  bibtex_key: stoller2018waveunet
  confidence: high
---

# wave-unet-musdb-eval

> Wave-U-Net: A Multi-Scale Neural Network for End-to-End Audio Source Separation — Stoller et al. (2018) (arXiv:1806.03185, 2018)

## What this evaluates

Evaluates end-to-end time-domain audio source separation models on singing voice and multi-instrument separation tasks. Probes the model's ability to isolate specific audio sources from mixed recordings using raw waveform inputs.

## Datasets

- **MUSDB** — total 150; splits: train (75), val (25), test (50)
- **CCMixter** — total ?; splits: train (-1)

## Metrics

- `MSE` **(primary)** — range: other
  - Mean Squared Error computed over all source output samples in a batch: (1/N) * Σ(y_pred - y_true)². Used for validation and early stopping.

## Input / output format

**Input**: Raw audio waveforms (mono or stereo), downsampled to 22050 Hz. Inputs are padded to match model context length. Mixture is formed by summing source signals.

**Output**: Separated time-domain audio waveforms for each target source.

## Scoring recipe

```python
def compute_mse(predictions, targets):
    # predictions and targets are 1D arrays of audio samples
    return np.mean((predictions - targets) ** 2)
```

## Common pitfalls

- Boundary artifacts can occur if input context/padding is not handled correctly during inference.
- SDR metrics are sensitive to outliers, which can skew performance rankings; rank-based statistics are recommended instead.
- Comparisons with prior work may be confounded by unknown or differing training hyperparameters (e.g., learning rates).

## Evidence (verbatim from paper)

> Final performance is evaluated on the MUSDB test partition comprised of 50 songs. As loss, we use the mean squared error (MSE) over all source output samples in a batch. We define 2000 iterations as one epoch, and perform early stopping after 20 epochs of no improvement on the validation set, measured by the MSE loss.

## Citation

```bibtex
@misc{stoller2018waveunet,
  title={Wave-U-Net: A Multi-Scale Neural Network for End-to-End Audio Source Separation},
  author={Stoller et al. (2018)},
  year={2018},
  note={arXiv:1806.03185}
}
```

- arXiv: 1806.03185