THE SHORT ANSWER
THE SHORT VERSION
- One request dubs a locked episode into up to 29 target languages; each language's files arrive by webhook as they finish.
- The review step pauses the dub after translation, so the last line before the cut is read before anything renders.
- A Do Not Translate list and a context brief set at the creator scope, which a studio uses for a whole series, apply to every episode; a job-scope entry overrides for one upload.
- Fast cuts, text over faces, and handheld footage are harder for Familiar's Alpha; a pilot on connected episodes decides which series fit today.
01.What does a daily release do to dubbing?
- One episode a day, every language the same day. The dub gets the hours between picture lock and the drop; a recording session per language does not fit that window.
- Drift compounds. A name or honorific rendered one way in episode 3 and another in episode 12 breaks a run viewers binge in one sitting.
- The last line is the retention beat. A translation that lands late or flat on the cut weakens the tap-through to the next episode.
- Everyone on camera is dubbed in their own voice; the score and the room stay under the new lines.
02.How does the same-day loop run?
| Step | What happens | What fires |
|---|---|---|
| Upload | The locked episode goes in once with up to 29 target languages; the source title and description ride along for translation. | A dub_id, one status per language |
| Review | The dub pauses after translation; the producer edits any line, the last one before the cut above all. | dub.ready_for_review |
| Render | One call approves the whole episode, or one language at a time while another language's reviewer finishes. | Per-language status: rendering |
| Deliver | Each language lands as its own file set: the finished video, .srt and .vtt subtitles, the translated title and description, the audio as the full mix or the voice alone. | dub.output_ready per language; dub.completed when all are in |
- Webhooks carry the schedule. dub.ready_for_review, dub.output_ready, and dub.completed report every state change, so a publishing queue moves without polling (Webhooks).
- Audio-only mode exists for audio cuts and voice tracks (dub_mode=audio); video with lip-sync is the default.
- 30 languages, any to any, in three published quality tiers, best first (the language table). Tier 1:
03.How does continuity hold across sixty episodes?
The voice comes from the clip itself, never a stock narrator, and there is no casting step: each episode's dub is generated from that episode's performance, so the actor carries their own voice into every language.
- One creator profile per series. Do Not Translate entries at the creator scope are persistent defaults: names, places, honorifics, and catchphrases entered once apply to every episode's dub, subtitles, and translated metadata, matched case-insensitively (Do Not Translate).
- A one-sentence brief at the same scope. "A revenge drama; Lin is the heroine, Zhao her stepbrother" is read before every line, so a name stays a name and a running setup keeps its payoff (Context).
- Narrower wins. A job-scope entry overrides the creator-scope default for one episode: a disguise arc, a flashback, a guest character.
- Changes apply from the next dub. A job already translating keeps the list it started with.
04.How does the cliffhanger survive translation?
- The transcript arrives per language with word-level timing. Every line carries its start and end in seconds and its speaker; the translated text is the one editable field (Transcript & review).
- The last line is read against the picture. The timestamps put the producer on the final seconds before the cut; a late or flat line is one edit, then render.
- Jokes and idioms are re-landed in the target language; a Do Not Translate entry pins any phrase that must stay exactly as said (What Matters in Video Translation).
- The face performs the new line. Eyes, brow, cheeks, and head re-perform with the new words, so the reaction shot before the cut reads in the target language.
- Review can be switched off so an upload renders straight through; on a series, one read of the final line per episode catches drift before it compounds.
05.What was measured?
The ElevenLabs numbers below come from paired outputs on the same source clips; the two ElevenLabs studies are separate cohorts.
ElevenLabs had 270% more background-sound error: the laughter, the music, the ambience.
11.92 VS 3.22 DB · ELEVENLABS DUBBING V2 (ALPHA) · 111 PAIRED OUTPUTS · MANDARIN, SPANISH, JAPANESE · AUGUST 5, 2026Familiar sounds 27.8% more like the real speaker; all 11 languages favored Familiar.
0.4605 VS 0.3604 · ELEVENLABS DUBBING V1 · 418 PAIRED OUTPUTS · 11 LANGUAGES · AUGUST 5, 2026Familiar made 54.7% fewer important spoken translation errors (29 vs 64).
ELEVENLABS DUBBING V2 (ALPHA) · SAME COHORT AS THE 270% TILE · AUGUST 5, 2026of the first-pass expert professional's quality, averaged across French, Chinese, Hindi, and Indonesian, scored blind against expert reference edits: a match.
PRODUCTION TRANSLATION STUDY VS FIRST-PASS EXPERT PROFESSIONAL TRANSLATORS · AUGUST 202606.What to test before committing a season
- Connected episodes, not one scene. Three to five consecutive episodes with the recurring lead, a confrontation, and a cliffhanger cut; the series list and brief set once before the first upload.
- The hardest episode first. The fastest cuts, the most text over faces, the most handheld camera; it decides which series fit today.
- A name change mid-run. Change a Do Not Translate entry between two episodes and confirm the next dub, its subtitles, and its translated title carry it.
- A late edit. A changed cut is a new dub request; the creator-scope list and brief carry over, and the review step is where the changed lines are read.
- Count corrections at the review step. Edits per episode, and whether the same edit recurs, are the pilot's numbers (How to Run an AI Dubbing Pilot (2026)).
QUESTIONS
Can AI dubbing keep up with a daily micro-drama release?
Yes. Familiar takes one locked episode into up to 29 target languages in a single request, pauses once at a review step, and renders every language together; each language's files arrive by webhook as they finish.
How do character voices stay consistent across sixty episodes?
The voice comes from the actor's own performance in each episode; nothing is cast or picked from a bank. Because every episode's dub is generated from that episode's lines, the same actor sounds like the same person on day 1 and day 60.
How do we stop names and catchphrases drifting between episodes?
A Do Not Translate list at the creator scope, which a studio uses for a whole series, applies to every episode's dub, subtitles, and translated title and description; a one-sentence context brief at the same scope tells the translator who the characters are. A job-scope entry overrides both for one episode.
Does vertical video work?
Uploads dub from any of 30 languages in either orientation; a portrait 9:16 dub is on the demos page. What is harder for Familiar's Alpha is what vertical micro-dramas often stack: fast cuts, text over faces, handheld footage. A pilot on the series' hardest episode shows which titles fit today.
REFERENCES
- [1]Do Not Translate: the three scopes, matching, seeded defaults (API reference)
- [2]Context: the one-sentence brief per scope (API reference)
- [3]Transcript & review: edit lines, render per language (API reference)
- [4]Familiar vs ElevenLabs benchmark: full results, intervals, and listening examples
- [5]Vs expert human translators: the translation-quality study
- [6]Why the Dub Feels Off Abroad, and How Studios Fix It (the studio playbook on series continuity)
Dubbing is finally good. See the measurements, then talk with the team about your catalog.
FAMILIAR · THE LAUNCH FILM · 2:31