Part of the cinematic pipeline: entry point creative-studio
Indic languages (script, voice, lipsync, subtitles)
Owners: screenwriter (script), voice-director (voice), editor (lipsync fit, captions), continuity-supervisor (QA). Labels: F fact (sourced in references/sources.md), 3P third party, E estimate, U unverified. AI drafts are drafts: a qualified native speaker reviews every final. Not a substitute for lab, legal or regulatory review; claims in scripts come only from the approved claims list.
Do not duplicate; link to:
- Language notes, cultural checklist, subtitle general rules, dub list:
${CLAUDE_PLUGIN_ROOT}/skills/film-studio/references/languages-and-localization.md - Indian accents, honorifics, code-switching basics:
${CLAUDE_PLUGIN_ROOT}/skills/asian-human-reference/references/speech-and-voice.md - Voice routes, credits ladder, naturalness score, spoken-Tamil script rules:
${CLAUDE_PLUGIN_ROOT}/skills/voice-studio/SKILL.mdandreferences/script-for-voice.md,references/indian-language-tts-options.md - App and website UI strings in English, Tamil, Hindi:
~/.claude/skills/i18n-tamil-hindi/SKILL.md(this skill is for spoken and film text, not UI strings) - Captions, transcription:
${CLAUDE_PLUGIN_ROOT}/skills/edit-voice/SKILL.md; mixing:${CLAUDE_PLUGIN_ROOT}/skills/music-and-sound/SKILL.md
1. Router (open ONE reference per job)
| Need | Open |
|---|---|
| Tamil script, voice, dialect, sample lines | references/tamil.md |
| Telugu | references/telugu.md |
| Malayalam (hardest script rendering) | references/malayalam.md |
| Hindi, Hinglish, schwa deletion, nukta | references/hindi.md |
| Kannada | references/kannada.md |
| Indian English, lakh and crore, accent casting | references/indian-english.md |
| Lipsync, dubbing, timing, ffmpeg fit, sync QA | references/lipsync-and-dubbing.md |
| Subtitles, fonts, UTF-8, transliteration standards | references/subtitles-and-transliteration.md |
| Sources, dates, unverified list | references/sources.md |
2. Quick card per language (E unless marked; confirm with native reviewers)
| Tamil | Telugu | Malayalam | Hindi | Kannada | Indian English | |
|---|---|---|---|---|---|---|
| Script, block | Tamil U+0B80 | Telugu U+0C00 | Malayalam U+0D00 | Devanagari U+0900 | Kannada U+0C80 | Latin |
| Shaping risk | Medium (two-part vowels, sri) | High (subscripts) | Highest (chillu, conjuncts) | High (matra, reph, nukta) | High (ottakshara) | None |
| Signature sound | ழ zha, ற, ட vs த | Geminates, no native aspirates | ഴ zha, ള, chillus, geminates | Aspirates, ड़, schwa deletion | ಳ, short and long e/o | Retroflex t d, even rhythm |
| Register gap | Large: centamil vs pechu | Medium: grandhika vs spoken | Large: formal vs colloquial | Medium: shuddh, Hindustani, Urdu | Medium: written vs spoken | Small |
| Polite you | நீங்க | మీరు | നിങ്ങൾ | आप | ನೀವು | you, sir, ji |
| Mix | Tanglish | Tenglish | Malayalam-English | Hinglish | Kanglish | any |
| Dub length vs English | Long (E) | Long (E) | Long, fast delivery (E) | Medium (E) | Medium to long (E) | Same |
| TTS leaders | Bulbul, ElevenLabs v3, Google | ElevenLabs v3, Bulbul, Google | ElevenLabs v3, Bulbul, Google | ElevenLabs, Bulbul, Google | Bulbul, ElevenLabs v3, Google | Bulbul en-IN, Google, ElevenLabs |
| Tool status | Listed (F) | Listed (F) | Listed (F) | Listed (F) | Listed (F) | Listed (F) |
| Listed means the vendor docs name the language: ElevenLabs v3 (official list), Sarvam Bulbul (ta, te, ml, kn, hi, en-IN, F docs), Google Chirp 3 HD (all six, F docs). Quality is U until a native listener scores a test. Lipsync tools are audio-driven; none documents Indic quality (U). |
3. Pipeline: script to voice to lipsync
- Brief: language, region, character age, class, register, who listens (audience), where it plays (ad, film, social).
- Script (screenwriter): write in the target language and script, spoken register, 6 to 14 words per line, one breath each. Three columns per line: native script, readable romanisation, English gloss (film-studio rule). English words inside the line: decide Latin or native script per tool.
- Native review of the script (gate 1). Fix register, honorifics, dialect, taboo, product claims.
- Pronunciation sheet (voice-director): names, brand and English words, numbers as words, acronyms, risky sounds (ழ, ഴ, schwa, aspirates). Respell per engine; store in the dictionary.
- Test line (credits ladder rung 1, voice-studio): the hardest sentence on 2 or 3 engines, 2 takes each; the user's yes before any paid call. Native blind score.
- Takes: Tamil v3 with at most one audio tag, 3 takes before committing; other languages 3 takes as well. Pick, splice, clean (voice-studio humanize-post.md).
- Time fit (editor): compare duration with the shot; adapt the script, not the speed (see lipsync-and-dubbing.md section 3). Speed change only within 0.9 to 1.1.
- Lipsync: choose by shot type and tool; 5 second test with the real audio before the full clip; no paid run without the user's yes (lipsync-and-dubbing.md section 4).
- Sync QA at 25 fps, frame by frame; fix within one frame; contact sheet saved (lipsync-and-dubbing.md section 6).
- Subtitles: same script, shortened to cps limits, UTF-8, correct font, libass or browser burn-in, never PIL without raqm (subtitles-and-transliteration.md).
- Native review of the final picture and sound (gate 2), then continuity-supervisor scorecard, then creative-director sign-off.
- Delivery: cast.md notes engine, voice, settings, consent and AI label; consent-log.md; the review sign-off name and date.
4. Hard rules
- Spend: quote first; no generation while researching or building; no paid call without the user's yes.
- Consent: no real-person voice or face without written consent; label AI voices; follow voice-studio voice-ethics.md.
- Mark untested language and tool combinations U; never claim a tool "supports" a language beyond what a source states.
- Never write English first and translate, unless the film is an adaptation.
- No accent as a joke; match accent to character and region, not to ethnicity.
- Do not strip ZWJ or ZWNJ; normalise NFC; UTF-8 for every text file.
- Subtitle and caption tools need HarfBuzz/raqm shaping. ffmpeg with libass and libharfbuzz is fine here; edit-director's Pillow has no raqm (checked 2026-10-14), so route Indic captions through libass or a browser renderer.
- Client-neutral: never reuse one client's script, voice or brand for another. No living public figures, film stars or politicians as names or voices.
- No TikTok for Indian clients.
- Rewrite, never rush: a line that does not fit is rewritten, then cut to a non-lip shot, before audio goes past 1.1x.
5. QA checklist for the native-speaker reviewer
Give the reviewer the script with the three columns, the audio, and the final cut. Ask for pass or fix on each item and record name and date.
- Meaning: the line says what the English gloss says; no machine-translation tone.
- Register: spoken, one register per character; no news-reader Tamil, no over-Sanskritised Hindi in casual scenes.
- Dialect: matches the place and character (Chennai, Madurai, Kongu, Jaffna; Telangana vs Andhra; Thrissur vs Kochi vs Malabar; Dharwad vs Mysore; Delhi vs Mumbai).
- Honorifics and address: age, gender, relation, community terms correct; verb agreement in Hindi.
- Sounds: Tamil and Malayalam zha, retroflex vs dental, aspirates (Hindi), schwa deletion, geminate length, vowel length; the brand name.
- Prosody: questions rise, statements fall, emotion fits, pauses carry meaning.
- Code-mixing: natural, phrase-edge switches, particles in small doses.
- Numbers, dates, currency, units spoken as a native would (lakh, crore).
- Sensitive content: religion, caste, regional politics, community stereotypes, sacred text; escalate to the Manager if unsure.
- Lipsync: mouth starts and stops with audio; closures p b m land; no teeth artefacts; still natural at normal speed.
- Subtitles: spelling, script shaping (no broken conjuncts), line breaks at phrase edges, cps within limits, correct font.
- Names and transliteration: one romanisation per name, native script recorded.
- Claims: product claims match the approved list; no testimonial voiced by a synthetic voice.
- Disclosure: AI voice labelled where required; consent log complete. Scoring: use the voice-studio naturalness checklist (8 or more ships) and the creative-studio scorecard (drafts 85, finals 90).
6. Known gaps (see references/sources.md)
No lipsync test has run for any Dravidian language; syllable rates for the five Indian languages are estimates; Higgsfield Translate language list not retrieved; all sample lines are own drafts needing native review; Azure and Cartesia current voices for Telugu, Malayalam and Kannada unconfirmed.