# Credits and measurement provenance

- Artificial inputs: fixed six-condition, four-input-seed experiment (24 inputs), generated by the supplied clock-event experiment code. No natural-speech recordings.
- Mean variation model: 平均変動対応版(v2), original saved candidate measurements.
- Coupling model: 相互作用対応版(v3), completed 24-encode / 96-decode run.
- PML and WORLD: original v1 saved measurements through the repository's vocoderbench.pml and vocoderbench.world interfaces. WORLD uses Harvest analysis. These are historical references, not new v3 measurements. Their original source/version hashes are retained in the compressed results.
- Experiment repository: Hiroshiba/check-nonml-saikyo-vocoder.
- Display tooling: Python, NumPy and Matplotlib. Audio files are PCM16 format copies of original float32 WAVs; no normalization, clipping, resynthesis or acoustic remeasurement.

The candidate has known source phase/F0 and common filter H, unlike the ordinary reference analysis. No natural-speech quality ranking or adoption conclusion is implied. PML/WORLD have no defined conditional Gaussian KL here. Original results and research payloads are not modified or republished as new measurements.

The named asset manifest identifies each audio file by condition, input seed, method, source version, target and decode seed. Exact source/public WAV hashes are in audio-provenance.json. Browser playback remains unverified.