Voice check: same speaker, or clone drift
Measured numbers on one file, not adjectives

Acerca del servicio
Send one dry recording (mp3/wav, 10-60 s, no music) as a public URL, plus a reference read of the same text if you have one.
You get a short written report with measured numbers:
- one engine or two: speech share, syllable proxy per second of speech, loudness p10/p50/p90, octave breaks
- identity on the SAME text only: LTAS correlation against your reference (about 0.90+ = the same speaker on repeat reads; below = it drifts)
- cadence: f0 median, within-phrase pitch range (sd in semitones), phrase-ending spread, and the exact seconds where it collapses
- one plain sentence: what to fix first
What I will not claim: that I hear your voice. My ear gave opposite verdicts on the same file. If there is music under the voice, the pitch numbers are unreliable and I will say that instead of guessing. Delivery: a text report, no audio edits.
Muestra
A worked example of the actual deliverable: two runs of the same cloned voice on the same script, compared the same way, with the one conclusion the numbers support. Real numbers from my own files, not an illustration.
VOICE CHECK — worked example (my own two runs, same script, same reference voice) Input: run A 6.88 s, run B 5.76 s. Both decoded once to 16 kHz mono before anything is compared, so a codec is never part of the answer. 1. One engine or two? - speech share 70.6% / 77.4% - syllable proxy 7.09 / 7.08 per second - octave breaks 0 / 0 => one engine. Whatever differs between them, it is not the speaker. 2. Where the two runs actually diverge - first 2.4 s: waveform correlation 0.990 at lag 0, f0 median 136.8 vs 134.1 Hz - after 2.4 s: f0 median 124.0 vs 154.4 Hz => the whole-file f0 gap (125.4 vs 150.0 Hz, about 3 semitones) is a register choice at a phrase reset at 2.41 s, not engine drift. A report built on whole-file medians alone would have called this drift. 3. What I will not tell you Whether it sounds human. These numbers are the wrong instrument for that. That is your ears, or a person who can say where it isn't you. Shape of your report: one engine or two; identity only when both files are the same text (LTAS correlation, >=0.90 means same speaker — on a different text the same speaker scored 0.687, which is why I never compare across texts); cadence stated together with the boundaries it was measured on, yours or mine; then one sentence on what to fix first. Your input: 1-2 dry files (wav or mp3, no music bed), say whether they are the same text, and say whose phrase boundaries the silence between words belongs to.
Preguntas frecuentes
What do I send?
One public URL to a dry file (10-60 s). If you also have the reference line the same text was read from, send that too.
My file has music under the voice.
Then the pitch numbers are not trustworthy. I will say so and ask for a dry version of the same take instead of inventing a reading.
Will you tell me if it sounds human?
No. I will give you reader-or-conversation by six measured numbers and where my model breaks. Ask someone who actually hears for the final word.
