After the script: Hinglish, voice and video

Turning the English script into Hinglish, building TTS prompts, word timing and the video plan.

1. Hinglish from a finished English script

From references/hinglish-adaptation.md. Do this after the English script reaches BAR MET. The English is the source of truth for facts, structure and order. The lint and the chain are English-anchored (quotes are English substrings, voice rules read English words), so they cannot judge the Hinglish. What can be checked deterministically is that the Hinglish kept the facts and the structure.

Steps

1. Write the Hinglish from the English, beat by beat, paragraph by paragraph: same beats and headings, same number of **VO:** paragraphs in the same order, same questions, same turns, same graphics lines (kept in English). Roman script, Hindi grammar, English words where Indian business speakers use them.
2. Run the parity check:

   node .claude/skills/video-script-production/scripts/hinglish-parity.js --english SCRIPT-vN.md --hinglish SCRIPT-vN-hinglish.md

   It must say PARITY OK: same beats and titles, same paragraph count, the same digits in each paragraph, the same number of question marks, every direct quote of four words or more verbatim, every capitalised name present, identical graphics lines. It reports the length ratio per beat. Moser Baer v7: 3,029 English words to 3,519 Hinglish (16% longer; the hook 18% longer).
3. Retime the headings with retime.js (150 wpm is an estimate).
4. Build the TTS prompts (section 2).
5. Record in the hand-off: the Hinglish has had no native read-aloud pass and no reader pass, its spoken length is unmeasured, and the lint does not apply to it.

Rules for the Hinglish itself (from the user's notes on the v6 Hinglish)

1. Use the word people say. Prefer a common English word when it is shorter and clearer: feeling (not khinchaav), busy, pivot, strategy, scale, healthy, signs, capital-intensive. Keep Hindi where it is the everyday word (tapri, kaudiyon ke bhaav, hadap liya).
2. Idiom, not translation. "ke just piche" for "just behind"; "haffti hui company" for a company panting on a treadmill; "dang reh jaate" for a double take. Replace any literal carry-over nobody would say.
3. One idea per sentence, the verb near its subject.
4. Do not gloss what the audience knows (crore, CBI).
5. Tag every person where they arrive and where they return ("Ratul Puri, Deepak ke bete aur company ke director", then "Ratul, Deepak ke bete" at each return).
6. Name a strategy once and call it back (scale; pivot, defined once in the hook, then counted first, second, third, fourth).
7. Say the turn: lekin, par, fir bhi, lekin phir at every BUT, but not three lekin in a row (use par).
8. Put the intensifier where someone pays: poore 17 saal, apni pay ka hissa tak.
9. The narrator has no stated gender. Use mujhe, mere liye, main only with verbs that carry no gender (dekha, nahi mila); avoid karunga/karungi, sakta/sakti.
10. Direct quotes stay in English; indirect speech stays indirect.
11. Numbers. Digits stay as digits (parity needs them). Numbers spelled out in the English are spelled out in the Hinglish (forty-two becomes bayalis). The TTS builder spells out currency and acronyms by rule.
12. Facts do not change in translation. If a note asks for a claim the sources do not support, write it in neither language and say so in the translation notes.

2. Gemini text-to-speech

Research and reasoning: research/the-fall-of-moser-baer/script/tts/TTS-NOTES-gemini.md. The user asked to try gemini-3.1-flash-tts.

The model situation

• Google's model page calls gemini-3.1-flash-tts-preview "a legacy preview model" with no shutdown date, and recommends gemini-3.8-flash-tts (best quality, 130+ languages) or gemini-3.8-flash-lite-tts (cheaper, 100+ languages).
• Tag syntax differs. 3.1: square-bracket English tags, about 200, including emotions ([awe], [determination]) and pace ([slow], [fast]). 3.8: the text is a verbatim transcript, style goes in a separate speech_metadata.style field, and only vocal or pause tags stay inline in angle brackets (<short pause>, <long pause>, <laugh>, <sigh>).
• Limits for both: 8,192 input tokens, 16,384 output tokens. One beat fits easily; the whole 20-minute script does not, so generate per beat.
• Output is 24 kHz mono 16-bit WAV.
• Hindi is supported. Hinglish and Roman-script input are not mentioned in anything read. Language is auto-detected; accent is set by the style prompt.

What is documented, what is third-party, what is unknown

The tool and its rules

node .claude/skills/video-script-production/scripts/tts-gemini.js --script SCRIPT-vN-hinglish.md --spec tts/spec-vN-hinglish.json --out tts --name vN-hinglish

It writes TTS-vN-hinglish-gemini-3.1.md (a direction block above the transcript, one fenced block per beat) and TTS-vN-hinglish-gemini-3.8.md (a style line separate from a verbatim transcript, only pause tags). The same works for the English script with spec-vN-english.json. The spec holds, per beat, a direction sentence, a style line, the tags ([snippet, tag]: the tag goes immediately before the first occurrence of the snippet after the previous tag) and the stress pairs (one word in capitals).

• It guards the words. It refuses to write unless stripping the tags and the capitals returns the normalised script word for word, every digit survives, no two tags touch, and every tag is on the documented list. Regenerate from the script and the spec; never hand-edit the TTS files.
• Normalisations by rule: "₹79 crore" becomes "79 crore rupaye" (rupees in English), "$150 million" becomes "150 million dollar", "CBI" becomes "C B I", "HR round" "H R round", "CD-R" "C D R", "VCDs" "V C Ds".
• Conventions used (design choices, not findings): about one tag per 55 words, mostly [short pause] and [slow]; [neutral] after a heavy passage; [curiosity] before each question that opens a loop; [positive] then [neutral] around the one quoted line with irony; stress capitals on a handful of words that carry the argument (TEZ, TAREEKH, MUSHKIL, WAHI, POORE; in English FULL, FASTER, DATE, SAME). The user placed [seriousness] before the first "Lekin" of the hook; keep placements the user marks.
• Reasoning for restraint: the channel voice is earnest about facts and dry about verdicts, which wants restraint more than acting. Moser Baer v7 Hinglish: 14 beats, 61 tags over 3,555 words, 12 stress words; English: 61 tags over 3,066 words.
• One blank line between voice-over paragraphs, never a break inside one. This is a guess (test T2).

The tests that settle what the docs cannot

Files that exist

• research/the-fall-of-moser-baer/script/SCRIPT-v7-hinglish.md and SCRIPT-v6-hinglish.md
• script/tts/: spec-v7-english.json, spec-v7-hinglish.json, TTS-v7-{english,hinglish}-gemini-{3.1,3.8}.md, TTS-v6-hinglish-gemini-{3.1,3.8}.md, TTS-NOTES-gemini.md
• script/audio/gemini-v7-hinglish/beats/B0-hook.wav to B13-close.wav (14 files), script/audio/beats/ (English v6 beats), narration-v7*.txt

Pace

Durations read from the WAV headers on 2026-10-27:

The Hinglish hook, 108.5 seconds, sits just inside the 110-second cap (the English v6 hook was 110 s at 150 wpm; the v7 English hook is 107 s). The per-beat Hinglish lengths run from 50.1 s (B2) to 110.1 s (B12). The Hinglish audio is almost the same length as the English headings (20:12) although it has 16% more words, so it is spoken about 16% faster than the 150 words per minute the headings assume (173.5 against 150). The plan flags this as a decision: at this pace the viewer has less room per number, and the pace is set by the TTS prompt, not by the timing code. The plan's own pace figure is an unfilled placeholder (PACE_WPM); the value above is derived (words divided by minutes) and is not in the plan.

3. Word timing

Plan and reasoning: research/the-fall-of-moser-baer/video/PRODUCTION-PLAN.md. The user wants a deterministic alignment, not a language model. A reference tool used an LLM for alignment; it was not copied. Whisper's text is not trusted; only its times are.

Whisper settings that work, tested on B2

Anchoring on B2 went 32% (first version) to 59% (LCS) to 86% (Urdu letters mapped). Final run: B1 285 of 315 words anchored (90.5%), B2 119 of 139 (85.6%); overall 404 of 454 (89.0%), 50 interpolated; 151.6 seconds of audio. One requested pause ("Agar", [short pause]) was checked: 0.23 seconds measured.

Limits

• Whisper base is the smallest multilingual model. Its words are poor; only its times are used. Word times are within about 50 to 100 ms. Interpolated words are estimates. Numbers written as digits are often spoken as words and so are interpolated.
• The global timeline puts beats end to end with a 0.5 second placeholder gap; the real gap is a production choice.
• Only B1 and B2 have been aligned; B0 and B3 to B13 have not.

Environment

• Python 3.8 venv at C:/Users/nnjha/.venvs/whisper38 (torch 2.4.1, transformers 4.46.3). Bash's python is Python 2.7 under Cygwin: use py -3.8 or the venv's python.exe.
• Run long jobs with -u, PYTHONUNBUFFERED=1 and PYTHONIOENCODING=utf-8 (Devanagari prints crash the Windows console otherwise). Do not put grep or a pipe in front of a monitored job.
• About 6 GB of RAM: Whisper base plus any second Python process hit "The paging file is too small" (os error 1455). Run one Whisper job at a time.
• Only openai/whisper-base is cached; small would need a download.

4. The video plan

Written 2026-10-25 for the Gemini Hinglish v7 beats. Stage 1 (timing) is built and measured; stages 2 to 5 are a plan, and nothing has been rendered or listened to by Claude. The plan's own assumptions, stated so you can correct them: "13 beats" is 14 WAV files (B0 plus B1 to B13); "Claude motion" is read as Remotion, the motion-graphics framework (its licence for this channel is unchecked); "being image search" is read as Bing image search, which Microsoft retired on 11 August 2025, so the plan does not use it.

Music and sound

• Recommendation: licensed library tracks for the beds, chosen by a rule; synthesised hits (ffmpeg) for the sound effects that need to be exact. Claude cannot listen, so it cannot judge an original score; AI-generated music depends on each tool's terms, which were not checked.
• Libraries named: Pixabay (no attribution needed per its FAQ), YouTube Audio Library (check each track), Freesound filtered to CC0. Content ID can still flag library music.
• The licence file is the rule: no track or image without a row in video/audio/licences.csv, no row without a source URL.
• Four acts from the beat directions: I the fall and the people (B0 to B2), II the rise and the price (B3 to B6), III the clock (B7 to B9), IV the people and the question (B10 to B13). The bed starts about 12 dB under the voice and ducks to about 18 dB under speech; measure after the first mix.
• Event rules come from the word flags in words.json: a soft tick on a number, a rising sting at the end of a question word, a low hit on a stress word (at most one per 20 seconds); no sound inside a long pause, because the silence is the effect; at most 8 sound events a minute and none within 0.3 s of another. No comedic stings, including in the tea-stall beat; none in B12, where the "alleged, not proved" label carries it.

Visuals: four categories and a rule for choosing

The rule for choosing, first match wins: a number, date, price, bar or line means A; a quotation, headline or source means B; a person, place or object with a licensed photo means C; a setting with no specific claim means D. The plan says not to re-typeset a newspaper headline as if it were a real clipping, not to use a news photograph without its own licence, not to show a photo of Mahesh Sharma or any private worker, and not to use the company logo without the user's decision.

Stitching (proposed, not built)

1. Timeline from words.json; clips placed by their anchor words.
2. Audio: the 14 beat WAVs as the voice; the act's music bed; ducking with ffmpeg sidechain compression; sound effects from the event rules.
3. Render with Remotion at 1920 by 1080, 30 fps; test on a two-minute preview (B0 to B1) first. Render time on this machine is unknown.
4. Loudness: two-pass ffmpeg loudnorm to about -14 LUFS.
5. Pre-render checks that can fail the render: every number on screen traces to a ledger row; every on-screen quotation is a capture or scan with the source's name and date; every picture and sound has a licence row.

Decisions the plan hands to the user

1. Pace: keep the delivered pace or regenerate slower?
2. Music route (licensed library recommended) and who auditions three candidates per act.
3. Source captures (category B): web page captures, scanned clippings, or only the ledger's own text on screen. Indian copyright law has a fair-dealing exception for criticism and review; this is not legal advice.
4. Stock footage with a "file footage" caption?
5. The Moser Baer logo or not (the default is no).
6. Whisper model: base is enough for timing; small would anchor more words but needs a download.
7. The real gap between beats (the timeline uses 0.5 s).

Not proven

The emotion-to-sound mapping is a hypothesis about how sound should follow feeling; no viewer has been tested. Claude has not heard any of the audio. The timing is only as good as Whisper base's timestamps. No licence has been checked for a specific track or image. Render time and the Remotion licence are unknown.

Written 2026-10-27 from the repository files and the project memory notes. If a page and a file disagree, the file is right: fix the page.