# Parallel M1 bakeoff audio sources

All six listening assets are continuous intervals from one original recording.
They are cropped and converted to mono PCM without denoising, dereverberation,
loudness normalisation, dynamic compression, concatenation, tempo change, or
pitch change. Model-native copies only add high-quality band-limited sample-rate
conversion.
Exact offsets and SHA-256 values are recorded in `manifest.json`.

## Natural meeting sources

Three distinct speakers are selected from the near-field headset recordings in
[AliMeeting Eval (OpenSLR SLR119)](https://www.openslr.org/119/). The corpus was
provided by Alibaba Group and the AliMeeting/M2MeT authors. It is licensed under
[CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/).

The preparation script parses the accompanying official TextGrid annotations.
It keeps 12–15-second wall-clock intervals that begin and end at official
utterance boundaries, with at least eight annotated seconds from one speaker,
no internal pause longer than 1.5 seconds, and zero annotated overlap with
another speaker. The three deterministic profiles are normal speech rate,
fastest eligible speech, and the highest expression-marker score, with different
speakers and recordings required for every profile.

## Clear-reading control source

`Zh-Google_penguin.ogg` was recorded and spoken by Wikimedia Commons user
[范博2004](https://commons.wikimedia.org/wiki/User:%E8%8C%83%E5%8D%9A2004).
The source page is
[File:Zh-Google_penguin.ogg](https://commons.wikimedia.org/wiki/File:Zh-Google_penguin.ogg),
licensed under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/).

## Fixed target references

- `Zh-前赤壁赋.ogg` was recorded and spoken by Wikimedia Commons user Ivier.
  [Source page](https://commons.wikimedia.org/wiki/File:Zh-前赤壁赋.ogg),
  [CC BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/).
- `Zh-Twitter.ogg` was recorded and spoken by Wikimedia Commons user
  [Wxzbb](https://commons.wikimedia.org/wiki/User:Wxzbb).
  [Source page](https://commons.wikimedia.org/wiki/File:Zh-Twitter.ogg),
  [CC BY 3.0](https://creativecommons.org/licenses/by/3.0/).

The targets are ordinary Wikimedia volunteers, not celebrity imitation
references. Attribution for the underlying spoken article text remains on each
linked Commons file-description page.

## Reproduction

```bash
.venv/bin/python scripts/prepare_bakeoff_audio.py
.venv/bin/python scripts/prepare_bakeoff_audio.py --verify
```

The raw downloads, extracted corpus files, WAV derivatives, and model-native
resamples are local generated assets excluded from Git. The small manifest and
this attribution notice are the reproducibility record that remains in source
control. Any redistributed BY-SA listening derivative must remain available
under the same CC BY-SA 4.0 terms with this attribution and modification notice.