Part of the cinematic pipeline: entry point creative-studio
Voice studio (human-sounding voice, Tamil first)
Owner: creative-director (casting, direction), fullstack-engineer (ffmpeg, pipelines), qa-tester (naturalness score), tech-writer (consent and licence log). Labels: F fact (sourced in references/sources.md), E estimate, U unverified, 3P third party and changing fast. Not legal advice.
Do not duplicate, link to:
- ElevenLabs tool names, voice rules, credit gate:
${CLAUDE_PLUGIN_ROOT}/skills/film-studio/references/elevenlabs.md - Sound plan, stems, dubbing flow, loudness table:
${CLAUDE_PLUGIN_ROOT}/skills/film-studio/references/sound-music-voice.md - Mixing, ducking, loudnorm, Indian and Tamil music, AI music prompts:
${CLAUDE_PLUGIN_ROOT}/skills/music-and-sound/SKILL.md - Vocal craft (voice types, registers, vibrato versus gamaka, Tamil diction, technique vocabulary, never imitate a named singer):
${CLAUDE_PLUGIN_ROOT}/skills/music-and-sound/references/vocal-craft.md - Indian accents, honorifics, code-switching:
${CLAUDE_PLUGIN_ROOT}/skills/asian-human-reference/references/speech-and-voice.md - Indian-language script, phonology, dialect, lipsync and subtitles per language (Tamil, Telugu, Malayalam, Hindi, Kannada, Indian English):
${CLAUDE_PLUGIN_ROOT}/skills/indic-languages/SKILL.md
1. Why AI Tamil sounds robotic (diagnose first)
Ranked by how often it is the real cause (E, from the research in references/sources.md):
- Script is written Tamil (செந்தமிழ் / news register), not spoken Tamil (பேச்சுத் தமிழ்). Engines follow the text; a formal script gives a news-reader delivery (3P, vendor guides).
- One long, flat generation: no choice of takes, no phrase-level splicing.
- Over-stable settings, or a voice that is not native to Tamil (accent leaks, flat pitch).
- Dry, sterile output with no room tone, no breath, no mix context.
- Emotion left to the model instead of written into the lines and directed with a few tags.
- Pronunciation errors in brand, English and number words. Singing has an extra limit: ElevenLabs Music docs name English, Spanish, German and Japanese for vocals; Tamil is not documented (F), so treat Tamil vocals as U until a test passes (references/hybrid-and-human-artists.md).
2. Router (open ONE reference)
| Need | Open |
|---|---|
| ElevenLabs v3 or v4 settings, tags, pronunciation, Voice Design, cloning, STS, failure fixes | references/elevenlabs-advanced.md |
| Which tool for Tamil or another Indian language, licence, price | references/indian-language-tts-options.md |
| Writing the script, spoken Tamil examples, numbers, punctuation | references/script-for-voice.md |
| Post-processing, ffmpeg recipes, room tone, breaths | references/humanize-post.md |
| Guide performance, speech-to-speech, hiring Tamil artists, singing | references/hybrid-and-human-artists.md |
| Consent, personality rights, disclosure, celebrity rule | references/voice-ethics.md |
| Sources, dates, conflicts | references/sources.md |
3. Decision tree: which route?
Answer in order; stop at the first match.
- Is the voice a real person's (founder, actor, singer)? Yes: written consent first (voice-ethics.md). Then a clone made with that person, or that person records.
- Is it a hero line (brand film, TV or OTT ad, emotional scene, any singing)? Quality bar is broadcast. Budget available: human artist, final recorded (hybrid-and-human-artists.md). Budget tight: human guide performance plus speech-to-speech, then polish.
- Can a person perform the lines (staff, friend, an artist for a short session)? Yes: speech-to-speech (STS). It keeps the performer's timing, breath and emotion and swaps only the timbre, often the most natural AI route (F, ElevenLabs docs: keeps whispers, laughs, subtle cues; the multilingual STS model lists 29 languages and Tamil is U, so test first).
- Volume content (explainers, social, product pages, many languages)? Quality bar is good web. AI TTS with a spoken-Tamil script, take selection and post (sections 4 to 6).
- Need one consistent branded voice at scale? Consented clone (PVC, or a Sarvam or Fish clone) or a Voice Design voice, locked per brand.
| Quality bar | Low budget | Mid budget | High budget |
|---|---|---|---|
| Internal draft, animatic | AI TTS, no post | AI TTS plus post | n/a |
| Social, web, explainer | AI TTS, spoken script, post | STS from a guide take | Human artist |
| Ad, brand film, OTT, songs | STS from a strong guide | Human artist, light AI fix | Human artist, studio, director |
Always run the same short Tamil test line through 2 or 3 engines (ElevenLabs, Sarvam Bulbul, one more) before choosing; native listeners score blind.
4. Naturalness checklist (0 to 2 each, total out of 10)
| Item | 0 | 1 | 2 |
|---|---|---|---|
| Prosody | Flat or sing-song, wrong stress | Mostly fine, a few odd phrases | Natural rise and fall; question and statement tones right |
| Breath and pauses | None, or in wrong places | Some, slightly regular | Breath at clause edges; pauses carry meaning |
| Pace variation | Same speed throughout | Slight change | Quickens on excitement, slows on weight |
| Emotion | Neutral or exaggerated | Right mood, shallow | Mood shifts inside the line, believable |
| Pronunciation | Names or Tamil sounds (ழ, ற, ண) wrong | Minor slips | Native reviewer approves; brand names right |
| Artifacts | Glitches, metallic ring, pitch jumps, clipped ends | Rare, minor | None on headphones and phone |
| Gate: 8 or more ships; 6 to 7 re-take or post-fix; under 6 change route (script, voice, or human). Scorers: one native Tamil speaker plus one non-expert listener, blind to the engine. Log the score. |
5. Credits ladder (never skip a rung; quote the cost, wait for the user's yes)
- Rung 0, free: rewrite the script in spoken Tamil, native review, write pronunciation notes. No credits.
- Rung 1: ONE short test line (2 to 3 sentences, about 150 to 250 characters) holding the hardest content: a retroflex word, a number, a brand name, an emotional beat. Run on 2 or 3 voices or engines, 2 takes each.
- Rung 2: pick the best voice and settings by score. Lock voice ID, model and settings in the log. One more test on the second-hardest line.
- Rung 3: generate the full script per line (not per scene): 3 to 5 takes for hero lines, 1 to 2 for the rest. Splice, post, score.
- Rung 4: finals only after the user approves the scored cut. Re-generate nothing that already passed. Stop rule: if rung 1 scores under 6 on two engines, stop spending and switch to speech-to-speech or a human artist.
6. Pipeline in one view
Spoken script with short lines > test line > voice choice > per-line takes > select and splice > clean (EQ, de-ess, compress) > room tone and subtle reverb > breath and mouth detail > mix under music > loudness check > naturalness score > native QA > consent and disclosure notes in the delivery file.
7. Hard rules
- Spend: quote first, the user's yes first. No generation while researching or building.
- Consent: no clone, no sound-alike, no celebrity, no deceased voice without documented rights (voice-ethics.md). No fake testimonials.
- Disclosure: label synthetic voice where platform or law requires (India IT Rules amendment 2026, EU AI Act timing; see voice-ethics.md).
- Confidential scripts and unreleased product claims do not go to a free or unvetted service.
- Product claims in voice scripts come only from the approved claims list; regulatory review stays with qualified staff.
- Client-neutral: never reuse one client's voice, script or brand for another.
- Native Tamil review before any final. Mark untested languages and features U.
Field test 2026-10-13: spoken-Tamil script + v3 + one [softly] tag on native voice Thivi = 3 takes for 363 credits; owner moving to ElevenLabs Enterprise (2026-10-14) for API settings, PVC and licensing.