Media AI agent · Subtitles & dubbing
Localization & Dubbing AI Agent (Subtitles, SDH & Dub Scripts)
Draft subtitles, SDH and dub scripts from the locked cut — timed, style-checked and consistent across a series — with no voice generated or cloned, and a native-language reviewer signing the language off.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Language sign-off
- Native reviewer
- After launch
- Agent Care
What this agent does
Prepares the language pass, never the voice
In
- 01
Take the locked cut, the as-broadcast audio and the house style guide for each language ordered.
- 02
Read the series glossary, the character names, prior episodes and the consents held on any voice in scope.
Reason
- 03
Transcribe the dialogue, time it to the shot changes and mark the on-screen text that has to be covered.
- 04
Draft subtitles and SDH inside the reading-rate, line-length and duration limits the style guide sets.
- 05
Adapt the dub script for length and lip sync, keeping names and recurring terms as the glossary has them.
Decide
- 06
Hold any line whose meaning turns on something the audio alone does not carry.
- 07
Route any generated or cloned voice to the person who holds that performer's consent.
Out
- 08
Hand the localisation supervisor a timed file with the reading rates, the glossary hits and the queries.
- 09
Retain the transcript, the style version, the queries raised and what the native reviewer rewrote.
Product statement
The agent transcribes, times and drafts. It creates no synthetic or cloned voice, and no language is delivered until a native-language reviewer and your localisation supervisor have signed it off.
Example workflow
One language pass, end to end
- 1Title and cut receivedThe locked picture, the as-broadcast audio, the languages ordered and the delivery spec
- 2Reference gatheredHouse style guide, series glossary and character names, prior episodes, and the consent held on any voice
- 3Transcript and timing builtDialogue transcribed and timed to the shot changes, with on-screen text and forced narrative marked
- 4Draft written and checkedSubtitle, SDH and dub-script drafts read against reading rate, line length, duration and the glossary
No human action required
Stages 1 to 4 run without a person in the loop — transcription, timing and the first draft finish before anyone is asked to read a line. A line that needs a voice with no consent behind it ends that stretch.
Goes to the native reviewer to sign.
Stops with the localisation supervisor.
Native-language reviewer
Reads the draft against the picture, the style guide and the glossary, and marks what has to change.
Approve · Rewrite · Send to supervisor
Approved — handed back
- 6Signed off by the reviewerDelivered only after a native-language reviewer signs that language; the agent signs off nothing itself
- 7Outcome reviewedReviewer rewrites, reading-rate failures, glossary breaks and anything corrected after delivery, by language
Reviewer edits
Lines the native reviewer rewrites are counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Creating or approving a cloned or synthetic voice.
- Accepting a language with no native reviewer.
- Deciding what a line means when the picture is ambiguous.
- Altering a performer's voice, accent or delivery.
Automation boundary
Agent acts unaided
- ✓Transcribe the dialogue and time each line to the shot changes.
- ✓Draft subtitles and SDH against the style guide in force.
- ✓Adapt the dub script for length and lip sync.
- ✓Mark every reading-rate, glossary and consent gap that it finds.
Write actions run only inside the approval boundaries agreed during implementation. Making a voice is not one of them.
- Rewriting a slur, a religious or a political reference.
- Declaring a subtitle or SDH track conformant.
- Releasing a language master to a platform.
- Changing the style guide, glossary or reading-rate limits.
Example output
One line in the language pass, annotated
Everything the agent drafts is attached to the cut, the style guide and the consents it read.
Localisation output · single line
Illustrative example- Title
- Korean original, episode 3
- Ordered
- German subtitle and dub
- Line
- 00:41:12
- Handed to
- Supervisor — held
- Confidence
- 87%
- Held on
- Retake asks for a cloned voice
As received
The episode, the source language and what was ordered — nothing on this side is inferred.
Evidence used
- House style guide v4
- Series glossary, ep. 1–2
- Consent file: nothing held
Why it was held
The retake would rebuild the performer's voice. Consent to that is the performer's to give.
What the score decides
Confidence sets how much the reviewer re-checks, never whether a voice may be made.
Action
ApproveRewriteSend to supervisorValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every language ordered
From the delivery schedule
Work to the style guide in force
Use the house style guide, the reading-rate limits, the series glossary and the character names already agreed.
Start review from a timed draft
Transcription, timing and a first subtitle and dub-script pass arrive done, so the reviewer reads a file instead of building one.
Raise the query before the mix
A line turning on something the audio does not carry, a glossary break or a voice with nothing on file is raised while it is still cheap to change.
Build an evidence trail
Retain the transcript, the style-guide version, the glossary hits, the queries raised, the reviewer's edits and what was corrected after delivery — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Media & post
- Avid MediaCentral
- Adobe
- Frame.io review
- conform and QC
Localisation tooling
- Subtitle editors
- CAT and TMS
- Translation memories
- glossaries
Assets & MAM
- Iconik
- Dalet
- As-broadcast audio
- scripts
Agent
Subtitling & dub scripting
- Times the transcript
- Drafts and checks
- Holds for review
Delivery & platforms
- IMSC / TTML
- SRT and EBU-TT
- Platform specs
- forced narrative
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and the delivered file
Each control wraps the one inside it. A draft clears every layer before a reviewer signs the language off, and making a voice sits outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide what the draft may be built from
- L3
- Measures the draft against the guide
- L4 – L5
- Keep sign-off with a person and log it
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the timing, the meaning and what the reviewer changed.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the timed file the reviewer opens
- E1Final-output evaluationDid the line carry the meaning the scene turns on?
- E2Step-level evaluationWas the transcript right, and timed to the shot change?
- E3Tool evaluationDid it read the current cut, style guide and glossary version?
- E4Reading-rate conformanceDid the draft sit inside the rate, length and duration limits?
- E5Slice evaluationHow does the draft hold across languages, genres and formats?
- E6Business outcomeHow much was rewritten in review, and corrected after delivery?
Floor — the lines the native reviewer had to rewrite
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle. A wrong line does not read as wrong to anyone who only speaks the language it went out in.
Agent lifecycle
Direction of processing →
01 · Source & reference
2 modes
- LD-01
Style guide out of date
Drafted to limits the platform has since changed.
- LD-02
Cloned voice with no consent
A supplied ADR track arrives with nothing on file.
Stage gathers
The cut, the audio, the style guide and the glossary
02 · Transcription
1 mode
- LD-03
Subtitle crosses the shot change
The line runs over the cut into the next scene.
Stage times
The dialogue, the shot changes and the timings
03 · Draft
2 modes
- LD-04
Plot point changed in the line
An idiom is flattened and the scene reads wrong.
- LD-05
Subtitle outruns the reading rate
It leaves the screen faster than it can be read.
Stage writes
The subtitle, the SDH and the dub-script lines
04 · Handover
1 mode
- LD-06
SDH cue and sign not delivered
A cue and an on-screen sign miss the track.
Stage presents
The timed file the reviewer reads and signs off
05 · Change / Version
1 mode
- LD-07
Silent glossary regression
A terminology update renames a character mid-season.
Stage tracks
Model, prompt, style-guide and glossary changes
- Sev-1 · a voice is used without consent
- Sev-2 · the viewer is told the wrong thing
- Sev-3 · the draft degrades and more is rewritten
Affected slices
The languages with the least data carry the rewrites
A pass that reads well in German can fail in the languages with the least data behind it, and in audio nobody scripted. They carry most of what a native reviewer rewrites. Nestack reports performance by slice, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Low-resource target languages | 6.3% | 3.6× | Review | |
| Archive audio, no script | 4.6% | 2.6× | Review | |
| Honorific and formality languages | 3.3% | 1.9× | Watch | |
| Scripted drama, script delivered | 1.2% | 0.7× | Normal |
Bar: reviewer-rewrite rate lift vs. scripted drama with a script · scale 0–4.0× · tick marks the 2.0× threshold
2 of 4 slices over threshold
Evidence-linked improvement
A line nobody in the room reads is the one that ships
The errors that survive are the ones your team cannot see. So the cycle runs on what the native reviewer changed, not on what the draft scored.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Reviewer rewrites, reading-rate failures or a post-delivery correction cluster in one language.
- 02
Diagnose
Traced to the transcript, the timing, the glossary, the style version or the adaptation itself.
- 03
Improve
The glossary, style rule or prompt is corrected with your supervisor and that language's reviewer, and version-linked.
- 04
Verify
Re-run against held-out episodes in that language, including the lines corrected after delivery.
- 05
Learn
The rewritten line is kept beside the original, and the rule it broke is written into that language's style guide.
Learn → Detect
The return edge. Style guides, glossaries and consent records are re-read each cycle — a name settled last season may have been re-spelled since.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, references and consents, drafting, evaluation, delivery formats, then supervised passes and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Workflow discovery and automation-boundary definition.Week 1
- 02Language list, delivery specs and style guides.Week 1, Week 2
- 03Voice consent, credit and compensation.Week 1, Week 2
- 04Transcription and timing to shot changes.Week 2, Week 3
- 05Series glossary and character names.Week 2, Week 3
- 06Subtitle and SDH drafting rules.Week 3
- 07Dub-script adaptation for length and lip sync.Week 3, Week 4
- 08Forced narrative and on-screen-text handling.Week 4
- 09Native-reviewer and supervisor review.Week 4, Week 5
- 10Evaluation suite, slices and regression episodes.Week 4, Week 5
- 11Delivery-format and platform integration.Week 5, Week 6
- 12Observability, deployment and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne title, one language | ProductionProduction media integration | AdvancedMulti-title / many languages |
|---|---|---|---|
| Introduced at Pilot | |||
| Transcription, timing and subtitle drafts | ✓ | ✓ | ✓ |
| SDH and forced-narrative drafting | ✓ | ✓ | ✓ |
| No synthetic or cloned voice created | ✓ | ✓ | ✓ |
| Native-language reviewer signs the language off | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Dub-script adaptation for length and lip sync | ✓ | ✓ | ✓ |
| Series terminology consistency | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Delivery-format and platform integration | — | ✓ | ✓ |
| Review workflow and observability | — | ✓ | ✓ |
| Additional languages and territory variants | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-title and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on languages and titles in scope, the subtitle, SDH and dub deliverables, media and delivery-system integrations, runtime volume, review controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01The locked cut, the as-broadcast audio and your delivery spec
→Language list, delivery specs and style guides
Week 102Voice consents, credits and what each one actually covers
→Voice consent, credit and compensation
Week 103Series glossaries, character names and prior episodes
→Series glossary and character names
Week 204Your house style guide and the reading-rate limits per platform
→Subtitle and SDH drafting rules
Week 305Episodes a reviewer sent back, and the lines they rewrote
→Evaluation suite, slices and regression episodes
Week 406The delivery formats and platform specs each language ships in
→Delivery-format and platform integration
Week 507Named localisation supervisors and native reviewers
→Review workflow, then supervised language passes
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Phases are drawn over the weeks they actually occupy. Week 5 carries both the delivery formats and the first languages a reviewer signs off.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Languages, deliverables and who reviews each one
W2Style guides, consent records and glossaries wired in
W3Transcription, timing and the subtitle and SDH drafts
W4Dub-script adaptation, forced narrative and the evaluation suite
W5Delivery formats and the first languages under review
W6Passes running under your supervisor, then Agent Care starts
- Reading the band
- Consent records and style guides are in place by the end of week 2, before a single line is drafted in week 3.
- At the end of W6
- Languages have gone out under a native reviewer and your localisation supervisor, with the transcript, the style version and every edit behind each one on file, then Agent Care takes over monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Media AI agent
Build a localisation agent around your style guide and your reviewers.
Show us an episode you have already delivered, the style guide and glossary behind it, and the languages you order. We'll draft one of those languages against your own guide and show you the lines it refused to decide and the voice work it would not touch.
- Typical delivery
- 4–6 weeks
- Deployment
- Your media and localisation stack
- Language sign-off
- Native reviewer, always
- After launch
- Agent Care available
Nestack Agents · Localization & dubbing
AGT-ME-07 · Agent Care available after launch