Industries / Media, arts and recreation / Media and entertainment / Film production
Editing and sound
Editing and sound is the work done after filming: cutting the footage into a story, setting the color and brightness of each shot, and building, mixing and translating the soundtrack.
7
Milestones achieved, from 1924 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1924
The first editing machine
Before the Moviola, a film editor, then called a cutter, held film up to the light and studied it frame by frame through a magnifier, according to a 1935 trade article. Iwan Serrurier's Moviola let the editor watch the film moving while cutting it; Moviola number one sold to Douglas Fairbanks on 16 September 1924 for $125. The Science Museum Group calls it the first machine for motion picture editing.
Credit
Iwan Serrurier, Moviola Company, Hollywood
1971
Disk editing arrives too early
CMX, a joint venture of CBS and Memorex, built the CMX 600, which TV Technology calls the first disk-based editing system. It was nonlinear: footage was stored so that any scene could be called up directly, with no winding through tape. Its picture was black and white and fit only for a first cut, it cost hundreds of thousands of dollars, and only a few were made.
LessonThe idea was right but price and picture quality were not ready, and TV Technology says it was more than a decade before nonlinear editing came back as a product.
Credit
CMX (CBS and Memorex)
1977
Stereo sound on ordinary prints
Dolby Stereo put a four-channel optical soundtrack on standard 35mm prints. Star Wars was not the first film in the format, which CBS News gives to Lisztomania (1975), but it made the system popular with the public and an industry standard.
Credit
Dolby Laboratories; the sound team of Star Wars
1989
Editing moves onto computers
Avid launched its first Media Composer system in 1989, sold as a package with its own video card and hard drives, and Lightworks, designed for feature film editors, was also first released in 1989. Early systems like these were "offline" tools, used to make the edit decisions but not the finished picture. In 1998 the Academy gave Avid a Scientific and Technical Award for its Film Composer system for editing motion pictures.
Credit
Avid Technology; Lightworks
2000
A whole film graded digitally
O Brother, Where Art Thou? was finished through a digital intermediate: after the edit was locked, Cinesite scanned the negative into a computer, and the cinematographer Roger Deakins spent about 10 weeks in Cinesite's digital suite with the colorist Julias Friede turning green trees and grass into dry browns and yellows before the result was recorded back to film. American Cinematographer notes that Pleasantville had already been digitized from end to end; on O Brother the whole film went digital to set its look.
LessonThe harder the image was pushed, the more noise and artifacts appeared, so Deakins kept changes to a minimum.
Credit
Roger Deakins; Joel and Ethan Coen; Cinesite
2012
Sounds placed around the room
Brave opened in Dolby Atmos in 14 theaters. Atmos sends each sound to particular speakers, including two rows of speakers overhead, instead of spreading it along the walls. The Associated Press called Brave the first feature mixed entirely for the format, and put the cost of equipping a theater at $30,000 to $100,000.
Credit
Dolby Laboratories; Pixar; the sound designer Gary Rydstrom
2025
Machine dubbing meets audiences
In May 2025 the Swedish film UFO Sweden opened in AMC theaters in the US as Watch the Skies: the cast re-recorded their lines in English, and Flawless's TrueSync software reshaped their mouths to fit the English words, a process called visual dubbing, in what Deadline later called the first full-length feature in theaters to use it. In December 2025 Amazon quietly removed AI-generated English voice dubs from anime such as Banana Fish on Prime Video after viewers and voice actors ridiculed them. In October 2026 Flawless's UK company entered administration, a UK form of insolvency, and its technology was sold.
LessonThe dubs that were pulled used synthetic voices, which the National Association of Voice Actors called "AI slop," while Watch the Skies kept its actors' own voices.
Credit
Flawless and XYZ Films; Amazon Prime Video
Watch for
Rough cuts built overnight
Eddie AI says its Night Shift feature takes a whole day's footage overnight, logs it, syncs the cameras and builds rough cuts, the first full versions of an edit, in a tool that works with Premiere Pro, DaVinci Resolve and Final Cut Pro. Its chief executive showed rough cuts assembled from hours of footage in minutes; the reports published no test.
Achieved whenAn editor or outlet not paid by Eddie AI publishes a blind comparison of Night Shift cuts against a human assistant editor's cuts of the same footage, or a credited editor of a released film or series says it built the assembly.
Claimed by
Eddie AI (chief executive Shamir Allibhai), NAB Show, April 2026, reported in May 2026
Automatic dialogue edit prep
AiAudioSoftware says DX Assist 2.6.0, a companion app for Pro Tools, rebuilds the production sound from the picture editor's edit list, finds every alternate take of a line by what was said, and writes ADR cue sheets, lists of lines to be re-recorded in a studio. Its Fresh Ear module, in beta, "compares the script against the recorded dialogue word by word and can send problem lines straight to ADR." The company says it is in use on broadcast series and features.
Achieved whenA dialogue editor credited on a released film or series describes using it in public, or an outside test reports how many of Fresh Ear's flagged lines match a supervising sound editor's ADR list for the same reels.
Claimed by
AiAudioSoftware, October 2026
Open
01
Score AI scene assemblies blind
A blind test on scripted scenes. Take 10 scenes with full coverage (every take and camera angle) and their script pages. AI tools build an assembly of each scene, the first full cut in script order, and three human assistant editors build their own from the same material. Working editors then watch every version without labels and score it.
The briefCredit · unclaimed02
Native speakers score visual dubs
A public test of visual dubbing. Take 30 short dialogue shots from openly licensed live-action films, have native-speaker voice actors dub each into three languages, and make three versions of each: the original-language clip as a control, the dub alone, and the dub with the mouths re-synced by an open lip-sync model. Native speakers of each language rate how well mouth and words match without knowing which version they see. Altered clips are labeled as test material.
The briefCredit · unclaimed
Open 01 · the brief
Score AI scene assemblies blind
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A blind test on scripted scenes. Take 10 scenes with full coverage (every take and camera angle) and their script pages. AI tools build an assembly of each scene, the first full cut in script order, and three human assistant editors build their own from the same material. Working editors then watch every version without labels and score it.
Why it is still open
Tools now claim to build rough cuts, but the published comparisons are run by the vendors on interviews, branded stories and creator videos. Wideframe had one editor cut four such projects both ways and two outside editors rate them; Editvideo.io (August 2026) asked 315 viewers, not editors, to compare a human edit with an AI edit of one creator video. This research found no published test on scripted scenes with multiple takes.
Built with what exists
Practice footage from EditStock, which sells unedited footage of films with all the takes and the script; Eddie AI's Night Shift and Quickture, a panel inside Premiere Pro and Avid Media Composer; a third assembler built on Claude Fable 5.2 (see the Note "Three working devices") that reads transcripts from ElevenLabs Scribe v3 (see the Note "Fewer word errors") and writes an edit decision list the editing software can import; Gemini Decision 1 as a machine grader to compare with the editors (see the Note "A thousand questions, one pass").
First step
Buy one EditStock project, pick two dialogue scenes, run one AI tool's free trial and one assistant editor on them, and try the blind scoring form on two editors.
Done looks like
10 scenes, each cut by at least two AI tools and three assistant editors; at least five working editors score every version blind on one 1-to-10 rubric covering story clarity, choice of takes, pacing and continuity; a published table of mean scores, time to first cut, and how often each version was ranked first.
Open 02 · the brief
Native speakers score visual dubs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test of visual dubbing. Take 30 short dialogue shots from openly licensed live-action films, have native-speaker voice actors dub each into three languages, and make three versions of each: the original-language clip as a control, the dub alone, and the dub with the mouths re-synced by an open lip-sync model. Native speakers of each language rate how well mouth and words match without knowing which version they see. Altered clips are labeled as test material.
Why it is still open
Research mostly scores lip sync with SyncNet numbers, which the THEval paper reports correlate poorly with human preference. The IWSLT 2024 dubbing evaluation used five clips in one language pair, English to Chinese, and the Face-Dubbing++ authors had to collect their own test set because none existed. The main commercial visual-dubbing firm, Flawless, sold its technology in administration in October 2026.
Built with what exists
Tears of Steel, a Blender Foundation short with real actors under a Creative Commons Attribution licence; LatentSync 1.6, ByteDance's open lip-sync model (Apache 2.0 licence, June 2025), which also supplies a SyncNet checkpoint for a machine score; ElevenLabs Scribe v3 to confirm each dubbed line was spoken as written (see the Note "Fewer word errors"); Gemini Decision 1 to compare with the human ratings (see the Note "A thousand questions, one pass"); any survey tool.
First step
Cut ten close-up dialogue shots from Tears of Steel, record a Spanish dub with one native speaker, run LatentSync 1.6 on them, and try the rating form on five Spanish speakers.
Done looks like
30 clips in 3 languages and 3 versions, each rated by at least 20 native speakers per language on a five-point lip-match scale, with a published table of mean scores and 95% confidence intervals and the correlation between human scores and the SyncNet score.