BRENNEN

Industries / Finance and business services / Translation services

Subtitles and dubbing

Subtitlers write a film's or video's dialogue in another language as short timed lines on screen; dubbing replaces the voices with speech recorded, or now generated, in another language.

5

Milestones achieved, from 1932 to 2025.

1

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1932

    Subtitles etched onto film

    Sound films ended the old practice of printed title cards between scenes; by Henrik Gottlieb's account, the first known showing of a sound film with subtitles was The Jazz Singer in Paris on 26 January 1929, in French. In 1932 R. Hruska in Budapest and Oscar I. Ertnæs in Oslo patented a method that pressed heated type into a thin wax coat on each print and bleached away the exposed emulsion, leaving clear white letters. Labs holding the main patents, Filmtekst in Oslo, Ideal Film in Stockholm and Titra-Film in Paris, led European subtitling from 1933 to the mid-1950s. Subtitling cost a tenth to a twentieth as much as dubbing, by the historian Jan Ivarsson's figure, so smaller language areas such as the Netherlands and Scandinavia chose it.

    web.archive.org

    Credit

    R. Hruska; Oscar I. Ertnæs

  2. 1988

    Lasers write subtitles

    A computer steers a narrow laser beam that burns each subtitle into the film's emulsion without damaging the base, taking less than a second for a two-line subtitle and needing no type or printing plates. Denis Auboyer and Titra-Film in Paris developed the method, in commercial use since 1988; it costs less per copy than the chemical process but needs costly equipment.

    web.archive.org

    Credit

    Denis Auboyer; Titra-Film

  3. 2018

    A streamer's translator test ends

    In March 2017 Netflix launched Hermes, an online test to find and rank subtitle translators, which it billed as the first by a major content creator. A year later it closed the test to new applicants, and in October 2018 Allison Smith of Netflix's localization team said the company had left recruiting, testing and training subtitlers to the ten or so vendors it works with.

    LessonOwning the whole chain from recruiting translators onward proved more than a streaming company wanted to run, and the work went back to specialist vendors.

    slator.com

    Credit

    Netflix

  4. 2025

    A film dubbed by face

    Watch the Skies, a Swedish science-fiction film released at home as UFO Sweden, opened in AMC theaters across the US on 9 May 2025 in a "visually dubbed" English version. The original actors recorded their lines in English, and Flawless's TrueSync software changed their lip movements to match, so the film looks as if it were shot in English. Press accounts called it the first full-length theatrical film to use the method.

    whattowatch.comengadget.com

    Credit

    Flawless; XYZ Films; director Victor Danell

  5. 2025

    AI anime dubs pulled

    In late 2025 viewers found English and Latin American Spanish dubs labeled "AI beta" on Amazon Prime Video for anime series including Banana Fish and No Game, No Life Zero. After voice actors spoke out, Amazon removed them; Anime News Network found more dubs gone by 4 December 2025. Kadokawa said it had not approved an AI dub "in any form", and HIDIVE said it was "not aware in advance".

    LessonA machine dub released without the rights holders' approval and without a quality check was withdrawn within days.

    animenewsnetwork.com

    Credit

    Amazon Prime Video

Watch for

  1. Dubbed and lip-synced reels

    Meta says its free AI translations tool, launched on 19 August 2025, dubs a creator's Facebook or Instagram reel into another language in the creator's own voice and tone, and can change their lip movements to match. It began with English and Spanish, and Meta's page lists later additions including Hindi, Portuguese, Bengali, Tamil, Telugu, Marathi and Kannada. It is open to Facebook creators with 1,000 or more followers and to all public Instagram accounts; the page gives no accuracy figures.

    Achieved whenan outside party publishes measured translation errors and lip-sync ratings for a sample of translated reels in several languages.

    creators.facebook.com

    Claimed by

    Meta, August 2025

Open

  1. 01

    Check AI subtitles against rules

    A public test of automatic subtitle translation tools on openly licensed films. Each tool subtitles the same films into several languages, and a script checks every subtitle against common rules: at most two lines, 42 characters per line, a reading speed of no more than 21 characters per second, and a time on screen between 5/6 of a second and 7 seconds. Professional subtitlers then mark meaning errors without knowing which tool made each file.

    The brief
    Credit · unclaimed
  2. 02

    Blind test of lip-synced dubs

    A blind test of AI dubbing tools that also change the speaker's mouth to fit the new words. Each tool dubs the same close-up clips of an openly licensed live-action film from English into three languages, using stock voices rather than clones of the actors' voices. Native speakers rate each clip, without knowing the tool, for natural voice, kept meaning and whether mouth and words match, and an automatic lip-sync scorer gives a number for every clip.

    The brief
    Credit · unclaimed

Open 01 · the brief

Check AI subtitles against rules

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test of automatic subtitle translation tools on openly licensed films. Each tool subtitles the same films into several languages, and a script checks every subtitle against common rules: at most two lines, 42 characters per line, a reading speed of no more than 21 characters per second, and a time on screen between 5/6 of a second and 7 seconds. Professional subtitlers then mark meaning errors without knowing which tool made each file.

Why it is still open

The IWSLT 2023 subtitling task set these line and speed limits, but it tested research systems on TED talks, European Parliament TV, fitness videos and ITV series with automatic scores. A Hugging Face set updated in November 2026 holds clips of the Chinese animated film Havoc in Heaven with subtitles from five AI models in six languages for human rating, but not the end-to-end tools that people buy, and not timing. The closest test of tools people buy, a 2026 EAMT paper by Lara Shoana Schlüter, Ekaterina Lapshinova-Koltunski and Sylvia Jaki, ran eight minutes of one TV episode through HappyScribe, CapCut and Amberscript from English into German, checked reading speed and line length among other things, and found the subtitles below human standards. WMT's 2026 video subtitle task tested research systems built on small open models, scored by AI judges with human correction. This research found no public test of commercial subtitle tools across several films and languages, checked against timing rules and graded by professionals.

iwslt.orghuggingface.coaclanthology.orgaclanthology.org

Built with what exists

Blender's live-action short Tears of Steel (Creative Commons Attribution 3.0), whose download page links community subtitles in eight languages to compare against; the auto-translate features of tools such as YouTube Studio and dubbing platforms that export subtitles; ElevenLabs Scribe v3 to make a clean transcript (see the Note "Fewer word errors"); SubER, AppTek's open-source scorer for subtitle edits; Subtitle Edit, an open-source editor, for timing checks; Netflix's timed-text rules for line count and duration.

  1. First step

    Run Tears of Steel through two tools into Spanish and German, and count rule breaks with a short script, side by side with the community subtitle files.

  2. Done looks like

    3 openly licensed films, 6 languages and 4 tools; the share of subtitles that break each rule, by tool and language; meaning errors per 100 subtitles graded blind by two professional subtitlers, with their agreement; all files and grades published.

Open 02 · the brief

Blind test of lip-synced dubs

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A blind test of AI dubbing tools that also change the speaker's mouth to fit the new words. Each tool dubs the same close-up clips of an openly licensed live-action film from English into three languages, using stock voices rather than clones of the actors' voices. Native speakers rate each clip, without knowing the tool, for natural voice, kept meaning and whether mouth and words match, and an automatic lip-sync scorer gives a number for every clip.

Why it is still open

Toloka's VOX-DUB benchmark, launched in November 2025, had native speakers compare four commercial dubbing systems blind, but only by ear; its authors say video and lip-sync are planned for later versions. Perso Dubbing's State of AI Dubbing 2026 data gives usage counts, not quality ratings. CROON, a dubbing company, published a May 2026 test of its own service against ElevenLabs, HeyGen, Rask and YouTube's auto-dubbing in which Gemini, not people, ranked the clips, and lip-sync was not among the criteria.

toloka.aihuggingface.cogithub.com

Built with what exists

Tears of Steel (Creative Commons Attribution 3.0); dubbing tools that offer lip-sync, such as HeyGen's video translation and Meta's AI translations for reels; SyncNet, an open-source model by Joon Son Chung and Andrew Zisserman (2016) that measures how well lips and audio line up; an online rating form; volunteer native speakers.

  1. First step

    Cut 10 single-shot close-ups from Tears of Steel, dub them into Spanish with two tools, and run SyncNet on the originals and the dubs.

  2. Done looks like

    30 clips, 3 languages and 4 tools, each clip rated by 10 native speakers per language; mean scores with 95% confidence intervals; the correlation between SyncNet scores and people's lip-sync ratings; clips, labeled as AI-altered, and ratings published under the film's license.