voice-clone-skill Checked

The verifier checks 24 kHz mono PCM output, re-transcribes it with the locally cached MLX Whisper Large v3 Turbo model, checks that the expected final sentence is present at the end, and runs RELEASE_CHECK: the peak amplitude of the closing 20 ms must stay at or below --release-threshold (0.02 by default; 0 disables it). Transcription alone cannot prove the ending survived, because Whisper reconstructs a clipped ~습니다 as complete text — a file can pass END_CHECK and CONTENT_CHECK while the last phoneme is audibly cut. A healthy ending decays to roughly zero; speech still at full amplitude on the final sample was truncated.

It also runs TRAIL_CHECK: the near-silent tail must be at least --min-trail-ms (100 ms by default; 0 disables it). This is a separate failure mode from RELEASE_CHECK and the one listeners notice most. A file can decay to a peak of 0.000 and still sound cut off, because the speech stops and the file ends 20-40 ms later — the ear hears a phrase at 0.3 amplitude vanish instantly. Pass --min-trail-ms 0 only for a chunk that the pipeline follows with its own inter-chunk silence, which supplies the pause.

If content verification fails, inspect the transcript and retry that narration at most once. After the retry, preserve the candidate and report any remaining near-homophone, proper-name, number, or semantic mismatch instead of looping.

Read a CONTENT_CHECK failure before acting on it. Whisper writes Korean numerals as digits (스물다섯 명 becomes 25명, 팔십삼 퍼센트 becomes 83%), so a mismatch confined to numerals, spacing, or a near-homophone is an ASR artifact, not a defective take. What does require regeneration is a transcript that ends with words the target text does not contain — the model sometimes tacks an extra fragment onto the end of a generation, and that fragment is really in the audio.

When a chunk fails RELEASE_CHECK, or is transcribed as complete but its release still sounds clipped, regenerate it with sentence-final punctuation appended to the target text. Append " ." first, then "\n.", then "."; the model finishes the last phoneme and emits its own trailing silence, so nothing is trimmed, faded, or padded afterwards. Only punctuation is added — never change, add, or remove a spoken word. Generation is stochastic, so a variant that fails once can succeed on a retry; cycle the variants and retry rather than concluding the sentence cannot be fixed. Verify each candidate with RELEASE_CHECK and confirm by transcription that the ending is intact. This applies to every affected internal chunk, not only the narration's final sentence.

### Sentence boundaries inside a chunk

A chunk holding several sentences is generated in one call, and the model tends to rush every phrase ending inside it: the

pauses between sentences shrink and the endings lose their release. RELEASE_CHECK and TRAIL_CHECK only see the chunk's

last sample, so they cannot detect this. Measure it against the source text instead — count sentences, then count the

150 ms-or-longer silences whose preceding 20 ms sits at or below 0.05. One clean pause per sentence boundary is the target.

Do not judge internal boundaries by scanning every silence for a loud edge. Korean stops (ㄱ/ㄷ/ㅂ/ㅈ) produce 50-120 ms

closures mid-word with the vowel still ringing into them; counting those reports dozens of false cuts and will not match

what the listener hears.

To fix a chunk whose internal boundaries are rushed, separate its sentences with a blank line (\n\n) so the model

treats each as its own paragraph. That reliably lengthens the internal pauses — but it also makes the model hurry the

chunk's own closing sentence, and no trailing punctuation rescues it. So split the work:

• join every sentence except the last with \n\n, generate that as one part;
• generate the final sentence on its own with trailing punctuation, as above;
• concatenate the two parts with a short pause (about 100 ms).

This is the same "final sentence generated independently" split generate_chunked.py already applies at narration level,

applied one level down. Retry each part separately against its own criteria: the body needs one clean pause per internal

boundary, the closing part needs RELEASE_CHECK and TRAIL_CHECK.

Prefer that method: it leaves the waveform the model produced. Fall back to the guarded-sentence recovery only if punctuation repeatedly fails — generate the exact sentence followed by a short neutral sacrificial sentence in the same model call, locate the sacrificial sentence's first word with Whisper word timestamps, then find the overlapping low-energy interval with silencedetect and trim inside that pause while retaining some of its naturally generated silence. Require a real inter-sentence pause (roughly 80 ms or longer) anchored on the detected guard position: Whisper word timestamps can overlap or coincide, and picking merely “the last silence near the end”, or trusting the target's own word end, can delete the target chunk's final sentence. Never infer the boundary from total duration. Do not fade, append silence, rewrite the target sentence, or include the sacrificial sentence in the delivered file. Use a versioned filename and re-run transcription on every trimmed chunk and on the assembled narration.

Only when the user or delivery format