# Training Free Multi Step Audio Sep Eval

> Evaluates a training-free iterative inference method for audio source separation. It probes the model's ability to progressively refine noisy audio mixtures (speech or music) by optimizing blending ratios across multiple inference steps without retraining or architectural changes. Use when the user wants to benchmark on VCTK-DEMAND, DNS Challenge v3, MUSDB18-HQ, or asks about evaluating this task. Reports PESQ, UTMOS, uSDR.

- Skill: `qhjqhj00/training-free-multi-step-audio-sep-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/training-free-multi-step-audio-sep-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/training-free-multi-step-audio-sep-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/training-free-multi-step-audio-sep-eval

---


# training-free-multi-step-audio-sep-eval

> Training-Free Multi-Step Audio Source Separation — Zang et al. (2025) (arXiv:2505.19534, 2025)

## What this evaluates

Evaluates a training-free iterative inference method for audio source separation. It probes the model's ability to progressively refine noisy audio mixtures (speech or music) by optimizing blending ratios across multiple inference steps without retraining or architectural changes.

## Datasets

- **VCTK-DEMAND** — total 824; splits: test (824)
- **DNS Challenge v3** — total 600; splits: test (600)
- **MUSDB18-HQ** — total 50; splits: test (50)

## Metrics

- `PESQ` **(primary)** — range: [1, 4.5]
  - Perceptual Evaluation of Speech Quality. Measures perceived speech quality by comparing enhanced output against clean reference. Higher is better.
- `STOI` — range: [0, 1]
  - Short-Time Objective Intelligibility. Predicts speech intelligibility based on temporal envelope correlation between reference and enhanced signals.
- `SI-SNR` — range: other
  - Scale-Invariant Signal-to-Noise Ratio. Measures objective signal fidelity independent of scaling, computed as the ratio of target energy to residual error energy.
- `UTMOS` **(primary)** — range: [1, 5]
  - Neural mean opinion score predictor for speech quality. Outputs a score estimating subjective listening quality without reference signals.
- `DNSMOS-P.808` — range: [1, 5]
  - Predicts BAK (background), SIG (signal), OVRL (overall), and MOS scores for blind evaluation using a neural model trained on DNS Challenge data.
- `uSDR` **(primary)** — range: other
  - Utterance-level Signal-to-Distortion Ratio. Computes the mean SDR across all test songs for each instrument stem.
- `cSDR` — range: other
  - Chunk-level SDR. Calculates the median of per-song median SDR values, where each song's value is the median SDR of all 1-second segments.

## Input / output format

**Input**: Noisy audio mixture waveform (speech with environmental noise or mixed music stems).

**Output**: Enhanced/clean audio waveform generated by iteratively blending the original mixture with previous step outputs using optimized mixing ratios.

## Scoring recipe

```python
def evaluate(predictions, gold, task):
    if task == 'speech_intrusive':
        return {'PESQ': pesq(gold, predictions), 'STOI': stoi(gold, predictions), 'SI-SNR': sisnr(gold, predictions)}
    elif task == 'speech_blind':
        return {'UTMOS': utmos(predictions), 'DNSMOS': dnsmp808(predictions)}
    elif task == 'music_sep':
        uSDR_vals = [sdr(gold_stem, pred_stem) for song in songs for stem in stems]
        cSDR_vals = [median([sdr(song_stems) for seg in song]) for song in songs]
        return {'uSDR': mean(uSDR_vals), 'cSDR': median(cSDR_vals)}
```

## Common pitfalls

- Using the faster/imperfect metric estimators (e.g., approximated PESQ, 6-second chunks with 50% overlap for SDR) during inference search for final reporting instead of standard evaluation metrics.
- Assuming monotonic improvement across all inference steps and metrics; SI-SNR consistently declines, and some music stems (Drums, Other) peak at intermediate steps rather than step 20.
- Comparing against diffusion-based models trained with multi-step objectives without accounting for differences in training data, architecture, or inference step counts.

## Evidence (verbatim from paper)

> We evaluate our approach using both intrusive and blind metrics. For intrusive evaluation, we report Perceptual Evaluation of Speech Quality (PESQ), Short-Time Objective Intelligibility (STOI), and Scale-Invariant Signal-to-Noise Ratio (SI-SNR). ... For music source separation evaluation, we employ two standard signal-to-distortion ratio (SDR) metrics on stereo signals: Utterance-level SDR (uSDR), which computes the mean SDR across all test songs; and chunk-level SDR (cSDR), which calculates the median of per-song values, where each song’s value is the median SDR of all 1-second segments.

## Citation

```bibtex
@misc{zang2025trainingfree,
  title={Training-Free Multi-Step Audio Source Separation},
  author={Zang et al. (2025)},
  year={2025},
  note={arXiv:2505.19534}
}
```

- arXiv: 2505.19534