# Attribution and provenance

All 27 public WAV files are normalized derivatives. The benchmark converted
the source audio to 16 kHz mono PCM and applied
`loudnorm=I=-20:TP=-2:LRA=7`. No source creator or dataset publisher endorses
Speakmac, Wispr Flow, or this comparison.

## Svarah — 18 clips

- Work: *Svarah: Evaluating English ASR Systems on Indian Accents*.
- Creators: Tahir Javed, Sakshi Joshi, Vignesh Nagarajan, Sai Sundaresan,
  Janki Nawale, Abhigyan Raman, Kaushal Santosh Bhogale, Pratyush Kumar, and
  Mitesh M. Khapra.
- Official source: <https://huggingface.co/datasets/ai4bharat/Svarah>
- License: [Creative Commons Attribution 4.0](https://creativecommons.org/licenses/by/4.0/).
- Changes: selected excerpts were resampled and loudness-normalized as
  described above.

The fixed 2026-07-28 corpus acquired these rows through the public
`amrithanandini/svarah` mirror, which is recorded in `manifest.json`. On August
31, 2026, the official AI4Bharat dataset card and CC BY 4.0 declaration were
rechecked. Official row-level access remained gated, so the mirror-to-official
bit identity could not be independently rechecked in this release pass. The
bundle preserves that provenance limit instead of claiming a stronger match.

Suggested citation: Tahir Javed et al., “Svarah: Evaluating English ASR Systems
on Indian Accents,” INTERSPEECH 2023, pp. 5087–5091.

## LibriSpeech — 9 clips

- Work: *LibriSpeech: an ASR corpus based on public domain audio books*.
- Creators: Vassil Panayotov, Guoguo Chen, Daniel Povey, and Sanjeev Khudanpur.
- Official source: <https://www.openslr.org/12>
- Subset: `test-clean`.
- License: [Creative Commons Attribution 4.0](https://creativecommons.org/licenses/by/4.0/).
- Changes: selected excerpts were resampled and loudness-normalized as
  described above.

The LibriSpeech utterance identifiers are preserved in the filenames and
sample-level manifest entries.

## Excluded macOS System Voice clips — 10 clips

The historical experiment included 10 benchmark stress scripts rendered with
macOS System Voices. Section 2F of Apple's current
[macOS Software License Agreement](https://www.apple.com/legal/sla/docs/macOSTahoe.pdf)
limits System Voice output to personal, non-commercial projects and disallows
public or commercial recording, publishing, and redistribution.

Accordingly, this public bundle contains no controlled System Voice WAVs,
reference files, or Speakmac 5.5.0 outputs. The older 37-row transcript and
aggregate files remain as historical text evidence, clearly outside the public
audio reproducibility set. No ownership or redistribution claim is made for
the System Voice output.