Chapter 7 — Voices
Status: DRAFT 1 (2026-09-21), DESIGN v6 §4. Sources: production/phase3/AUDIO-PLAN.md rev 2; ledger rows cited by line.
7.0 Objective
Given the script, you can audition voice models on one line by ear, record every line as its own file with a tone tag, keep an audio ledger, and say what an attached audio reference does and does not promise.
7.1 What you are deciding
Every line in your film will be a separate audio file made by a text-to-speech tool, before any face is generated. Three decisions:
- Which voice, per character. Chosen by ear on one line, from two or three candidates, then never changed: same model, same preset, no setting touched between lines.
- One clip per line, with real silence. The card's silence before the line is silence in the file, so the video model later has nothing to guess.
- What the audio is for. In the edit it is the line. Attached to a generation it is a hint: it conditions, it does not promise timing.
Voice identity is settled here. Whether a delivery has intention is Chapter 8; whether a mouth matches it is Chapter 11.
7.2 How we did it
Audition (AUDIO-PLAN.md:22–25). Three text-to-speech models on Crystal's line 我耳环掉了。不知道滚哪儿去了。 at 0.01, 0.15 and 5.7 credits per take (ledger rows 12–13). Whisper transcribed the 0.15-credit take word for word, 3.12 s, −17.0 dB. The Director listened and chose: "eleven is the best" (AUDIO-PLAN.md:56); for Rock, "Holden is best" (row 31). Total audition spend under 10 credits.
Batch (AUDIO-PLAN.md:61–72). Twenty line files: eighteen new jobs, two reused from the audition; price read before each click (0.15) except one read after (1C, at 0.45, logged as a protocol miss); transcripts word for word on all twenty with one accepted contraction noted. Padded 1 s lead + 0.5 s tail, 44.1 kHz mono, 256 kbps MP3, named aud_HW_<card>_line_v1.mp3, uploaded into the platform project as the festival required (asset count 81 → 100). Total new spend 2.85 credits.
Tone tags. The model took a bracketed cue in front of the text. The fence for a line is the cue plus the text and nothing else; the whole fence for 2A was [casual] 是巧克力。 (production/phase3/post-submission/2A-audio-B-fence.txt). Cues that worked: [casual], [conversational], [playfully teasing], [quickly, casual, a little curious]. One that did not: [laughs] alone was rejected by the platform (row 527, job failed); [laughs] 哈哈哈 worked (row 529).
Hard limit (AUDIO-PLAN.md:29): "Never a real person's voice."
7.3 Case 7-A — the attached audio that did not set the timing
- What I saw: a 7 s generation for card 1B, sound off, with the approved line attached as an audio reference and its timing written in the prompt (speech at 4.36–5.01 s). The result "kept phone away during that interval and moved the mouth afterward" (LESSONS.md:136). 32 credits (rows 393–395).
- Why I think so: the reference conditioned the take (a mouth moved) but its timestamps were not followed; an earlier take had ignored the 1 s lead-in silence the same way (LESSONS.md:82–83). The vendor's wording "syncs… generates lip sync" does not promise that timestamps are honoured (1B-proof-review/CLAUDE-VERDICT.md:29).
- What I changed: the rule in DESIGN §6: an audio reference conditions, it does not promise verbatim playback or timing; the approved line is laid in the edit and accepted per the sync rule.
- Would this happen on your tool? Unknown, and that is the point: no controlled comparison exists (LESSONS.md:136 says so in those words). We did not run the same card with sound on, so "sound off caused it" is not established. On your tool, run the two arms on the cheapest duration before you rely on either.
7.4 Do it with your tools
| Need | We used | Any tool | If yours lacks it |
|---|---|---|---|
| Per-line TTS with a tone cue | a hosted v3 model in the platform, 0.15 per line | any TTS that takes a style or emotion cue | put the tone into the words (a shorter sentence, a comma) and audition by ear |
| Audition by ear on one line | three models, one line, Director listens | same | the second reader listens with you |
| Upload into the project | the platform's asset upload | wherever your destination's verifier looks | keep a folder with the tool named in the ledger |
| Transcript check | whisper | any offline ASR | listen twice |
7.5 Your turn
Time cap: one evening. Spend: from your Chapter 6 sheet's voice row; ours was under 10 credits for the audition and 2.85 for twenty lines.
- Pick the audition line: the hardest line in your script (longest, or the one with the turn).
- Audition two or three voices per character on that line. Fence each (cue + text), price read, one click, ledger row. Listen with the second reader, sound only, no picture. Ledger the choice as a decision with one sentence of why.
- Record every line as its own file: 1 s lead, 0.5 s tail, one cue in front. Name files by card. Transcribe each; a transcript that is not word for word is a retake (same voice, reworded cue or punctuation).
- Store every file where your destination requires and write the tool's name in the ledger row.
- Do not attach any of these to a video generation yet. Chapter 11 decides that per card.
7.6 Gate
Each line needs valid evidence per DESIGN §5; outcomes are pass / fail / unverified, and unverified blocks.
- One voice per character, chosen by ear on one line, with the decision row naming who listened (evidence: ledger; self-attested listening).
- Every line exists as its own file with lead and tail silence, transcript word for word (evidence: the files and transcripts).
- Every audio job has a fence, a queued row with the read price, and a read row (evidence: ledger).
- Performance task: find the sentence in your own tool's documentation about attaching audio to a video generation, paste it, and write the two-arm test (same card, cheapest duration, with and without the audio attached) that would tell you whether timing is honoured; run it only if your Chapter 6 sheet has the credits (evidence: the pasted sentence and the test design; the result if run).
- Files are stored where the destination's verifier can see them, tool named (evidence: the asset list or folder listing).
Pass example from HALFWAY: the line batch, eighteen jobs, price read before each click but one (logged), 2.85 credits, assets 81 → 100. Fail example: the 1B proof, 32 credits spent on timing the tool never promised, with no second arm to learn from.