Voice shootout, Stayfound card-cancel mash-up

The ad's own narrator line, word for word, in each voice. Red cells sit outside the pass bars (small print under each heading).

measure	phrase s 3.2-5.9	pause ms 190-225	fall st 2.0-2.7	spread st 5+	wpm 180-200	pitch Hz
this option	5.91	140	2.02	5.85	158	142.9
reference narrator	5.91	140	2.02	5.85	158	142.9

For comparison: the 2026-11-02 voice Kalp rejected

measure	phrase s 3.2-5.9	pause ms 190-225	fall st 2.0-2.7	spread st 5+	wpm 180-200	pitch Hz
this option	2.85	175	0.4	5.49	197	139.7
reference narrator	5.91	140	2.02	5.85	158	142.9

Model: Inworld TTS-2, narrator clone, comma joins, one take kept
Sample: DeepInfra voice moejqzjplghag5o2un27 (18 s sample)

Whisper hears "opened" where the script says "open"; 0% of its phrases end low.

Option 1: Inworld TTS-2, narrator clone

measure	phrase s 3.2-5.9	pause ms 190-225	fall st 2.0-2.7	spread st 5+	wpm 180-200	pitch Hz
this option	2.95	190	1.86	5.77	185	135.0
reference narrator	5.91	140	2.02	5.85	158	142.9

Model: Inworld TTS-2 (DeepInfra), clone of the Exposr news narrator
Sample: DeepInfra voice moejqzjplghag5o2un27: the 18 s Demucs narrator sample (beat the new 30 s narrator-only sample rolcnymp770mb8786kow on 6 takes each)

Kept take: o1-inworld2-s18-join-t3.mp3 of 12 takes (tempo 1.0). Word check: clean.

Option 2a: Xiaomi MiMo-V2.5-TTS voiceclone, narrator clone

measure	phrase s 3.2-5.9	pause ms 190-225	fall st 2.0-2.7	spread st 5+	wpm 180-200	pitch Hz
this option	6.25	160	0.69	5.9	193	139.1
reference narrator	5.91	140	2.02	5.85	158	142.9

Model: Xiaomi MiMo-V2.5-TTS voiceclone (DeepInfra), clone of the same narrator
Sample: DeepInfra voice rolcnymp770mb8786kow: 30 s narrator-only Demucs vocals

Kept take: o2-mimo-s30-join-t5.mp3 of 12 takes (tempo 1.079). Word check: clean.

Option 2b: Resemble Chatterbox, narrator clone

measure	phrase s 3.2-5.9	pause ms 190-225	fall st 2.0-2.7	spread st 5+	wpm 180-200	pitch Hz
this option	8.05	200	0.63	4.2	184	130.1
reference narrator	5.91	140	2.02	5.85	158	142.9

Model: Resemble Chatterbox (DeepInfra), clone of the same narrator, exaggeration 0.5, cfg 0.5
Sample: DeepInfra voice rolcnymp770mb8786kow: 30 s narrator-only Demucs vocals

Kept take: o2-chatterbox-s30-orig-t2.mp3 of 6 takes (tempo 1.0). Word check: replace: '18' -> 'one eight'.

Option 4: cloned real creator @contentbyroxy

measure	phrase s 3.2-5.9	pause ms 190-225	fall st 2.0-2.7	spread st 5+	wpm 180-200	pitch Hz
this option	3.06	180	3.93	3.62	199	192.8
reference narrator	5.91	140	2.02	5.85	158	142.9

Model: Inworld TTS-2 (DeepInfra), clone of TikTok creator @contentbyroxy (her clip is already in this ad)
Sample: DeepInfra voice 0ar6nqop278m0hkesfsb: her own TikTok 7638778345055882516, fetched on AdsPower research, Demucs vocals 0-32 s

Kept take: o4-roxy-inworld2-join-t2.mp3 of 12 takes (tempo 1.0). Word check: clean.

Her own voice, 15 s of the sample: spread 3.41 st, 217 wpm. Her clip already plays in this ad. A clone of her needs her consent before it runs.

Option 5: real clip audio for 30 of 52 words + TTS

measure	phrase s 3.2-5.9	pause ms 190-225	fall st 2.0-2.7	spread st 5+	wpm 180-200	pitch Hz
this option	5.36	190	0.56	5.23	171	115.1
reference narrator	5.91	140	2.02	5.85	158	142.9

Model: Real clip audio for 30 of 52 words (Demucs vocals of four ads in our library), Inworld TTS-2 narrator clone for the other 22, then tighten_vo
Sample: clip runs listed in plan; TTS gaps cut from take o1-inworld2-s18-join-t3

Kept take: o5-clips-plus-inworld2-tight.mp3 of 2 takes (tempo 1.0). Word check: replace: 'open' -> 'opened'.

Which words are clip audio:

Just 1.8 millimeters thin, (clip audio: Spotminders ad DLXWBf0gA7n (research/ads/top/02), 8.06-9.76 s)
it blends (TTS: o1-inworld2-s18-join-t3.raw.mp3)
right into your wallet. (clip audio: Spotminders ad 1129506182737235, 22.00-23.06 s)
Apple Find My compatible, (clip audio: Spotminders creator ad DJubqZYALin (research/ads/top/09), 32.41-33.99 s)
with a loud beeping sound for close range recovery, a rechargeable battery and no subscription fees. (TTS: o1-inworld2-s18-join-t3.raw.mp3)
So (clip audio: Spotminders ad 1106242948297304 (batch donor-03), 29.06-29.32 s)
I just open my phone and there's my (clip audio: same ad, 30.24-32.44 s (its words between, the moment I land, are left out))
wallet (TTS: o1-inworld2-s18-join-t3.raw.mp3)
right on the map. (clip audio: same ad, 32.78-34.64 s)
Buy two, get one free, (clip audio: same ad, 52.92-54.12 s)
with free shipping. (TTS: o1-inworld2-s18-join-t3.raw.mp3)

Bold = a real person in one of our library ads (Spotminders ads, 4 different speakers); the rest = option 1 voice. Running Spotminders audio in our ad needs clearing first.

Findings

Inworld TTS-2: the 18 s narrator sample beat the new 30 s narrator-only sample (best take 0.53 outside the bars against 3.75).
Higgs Audio V2.5 on DeepInfra: preset voice works, every cloned voice (12, 18, 30 s samples) returns inference error.
Inworld TTS-1.5-max smoke take: 117.6 Hz against the narrator 142.9 Hz, pauses 320 ms.
Chatterbox reads 1.8 as one eight on all 6 takes and makes nonsense syllables at the ellipses.
MiMo adds babble on some takes (Whisper: uh i think uh), so every MiMo take needs the word check; 3 of 12 MiMo takes were clean.
Roxy own voice measures 3.41 st spread at 217 wpm, so her clone at 3.6 st is flat by her own voice.
Whisper hears opened for open my on several takes and on option 5, where those words are a real person recording, so that flag over-fires on this join.
Option 5: none of the ad own creator clips says 3 or more exact words of the line; the 30 clip words come from 4 Spotminders ads in our library, 4 different speakers.
Rights: option 4 clones a creator who has not agreed (voice_cast.py keeps clones to consenting people) and option 5 plays audio from Spotminders own ads; neither ships without consent or our own recording.

Spend: $0.291 (46 TTS calls + Vast + Whisper).

Voice shootout: HR-STAYFOUND-BROLL-card-cancel-mashup-v01

Reference narrator (Exposr news narrator, Demucs vocals, 39 s)