m30-aterencehughes — Aaron Terence Hughes, perfumer and fragrance reviewer
clone pitch 87.7–94.7 Hz · sample 8,625 Hz → 12,000 Hz · clarity 26.5 → 24.4 dB · Whisper clean on 2 of 3 takes
old, 8.6 kHz sample
new, 12 kHz sample
You asked whether the text-to-speech model was the wrong one. It was the right one: DeepInfra's Inworld Realtime TTS-2, the same engine behind all thirteen voices you approved on 3 November. The defect was in the sample each clone was built from.
The script that found these voices downsampled its audio to 16 kHz before I cut the clone samples. A 16 kHz file cannot carry anything above 8 kHz, and that is where sibilance, breath and consonant air sit. Measured: the three samples I registered on 5 November carried 8,625, 8,554 and 8,203 Hz. Every voice you approved carries 10,851 to 12,000 Hz. A clone cannot invent what its sample never held.
I re-pulled the three sources keeping the original audio stream, cut new samples at 24 kHz, and registered three new clones. All three samples now measure 12,000 Hz, matching Dr George, the voice you called the most authoritative. Each voice below plays the old clone and the new one on the same two sentences of the Rainbow Passage.
| voice | sample bandwidth | pitch | pitch spread | voiced share | seconds | Whisper |
|---|---|---|---|---|---|---|
| aterencehughes old | 8,625 Hz | 84–100 Hz | 3.70–3.83 st | 0.23–0.27 | 8.2–8.8 | 1 word out in 3 takes |
| aterencehughes new | 12,000 Hz | 88–95 Hz | 3.37–4.17 st | 0.22–0.23 | 8.2–8.6 | 1 word out in 3 takes |
| balancegrooming old | 8,554 Hz | 90–92 Hz | 2.93–3.91 st | 0.22–0.25 | 8.2–8.4 | clean |
| balancegrooming new | 12,000 Hz | 91–94 Hz | 2.90–3.74 st | 0.19–0.24 | 7.8–8.6 | 2 words out in 3 takes |
| mattytingles old | 8,203 Hz | 93–103 Hz | 3.30–4.13 st | 0.59–0.69 | 8.9–10.2 | clean |
| mattytingles new | 12,000 Hz | 92–107 Hz | 3.49–3.85 st | 0.58–0.60 | 9.0–9.7 | clean |
| wolfwilliamsburg (approved) | 10,851 Hz | 120–126 Hz | 5.44–6.15 st | 0.11–0.13 | 8.1 | clean |
| drgeorge (approved) | 12,000 Hz | 110–128 Hz | 4.24–4.77 st | 0.15–0.20 | 8.2–8.3 | 2 of 3 takes misread a word |
I flagged the new voices as having a flat pitch spread, 3.5 to 4.5 semitones against 5.4 to 6.2 for an approved voice, and called that the robotic tell. I had measured against one approved voice only. Dr George, the one you picked out as the most authoritative, reads 4.24 to 4.77 semitones, which sits inside the range I called flat, and his source audio is flatter still at 2.98. Pitch spread did not separate the voices you approved from the ones you rejected, so I should not have leaned on it. Bandwidth does separate them, with no overlap at all.
The hunt script now keeps a copy of the source audio stream beside its 16 kHz measurement file, so a clone sample is never cut from a resampled file again. The clarity tool measured bandwidth all along and printed it beside its verdict without ever judging it, which is why all three samples read "good"; it now returns "narrow" under 10,000 Hz. The takes also went through the house caster this time, so the spend is in the ledger and the pace, pause, phrase-fall and Whisper checks ran, none of which I had measured on the first batch.
Your verdict is still empty on all three, and the Out There ad batch remains stopped where you left it on 5 November.