Built with Claude
Dubbing that knows who's talking.
The fastest way to ruin a dub is to give a line to the wrong voice. Monoglot uses Claude by Anthropic to decide which character speaks every line, and to explain each decision in writing.
Two Claude models
Sonnet by default. Haiku at volume.
Careful on every line
Handles off-screen speech, similar voices, crowded scenes, and listener shots. Sonnet weighs the frames, the dialogue, and the audio evidence together and explains each decision.
Fast, high-volume runs
Choose Haiku per project for large catalogs and quick social cuts, where speed and cost matter most.
Claude works from text and images. It never listens to audio or watches video directly. Our own audio and video models turn the scene into transcripts, voice-match scores, and still frames that Claude can read.
The hard part
Voices alone aren't enough.
Two characters with similar voices. A line spoken off screen. A reaction shot of the listener while someone else talks. Audio clustering gets these wrong, and every mistake becomes a wrong voice in the final dub. Claude weighs all the evidence together, the way a human editor would.
- line
- “You said you'd be back before dawn.”
- character
- Mira
- confidence
- 0.91
- needs review
- no
- reason
- Camera holds on Kaito, but his mouth is closed. The voice matches Mira's cluster, and the line answers Kaito's previous question. Off-screen speaker.
How Claude fits in
Specialist models see. Claude reasons.
- 01
Gather evidence
Our audio and video models separate and cluster voices, track faces, and watch who is on screen. Each line becomes text evidence plus a few still frames.
- 02
Claude decides
Claude reads the line, the dialogue around it, the voice candidates, the character roster, and the still frames. It returns a character, a confidence score, and a written reason.
- 03
Guardrails hold
Claude can only choose characters that exist in your project. Off-screen speech, listener shots, and weak evidence are flagged. Anything below the confidence threshold is never applied automatically.
- 04
You stay in control
Studio mode shows every decision next to Claude's reasoning. Approve it, override it, or send it to a reviewer before a single voice is generated.
For studios, platforms, and localization teams
Claude-checked dubbing at catalog scale.
Localization teams can't listen to every line of every episode. Monoglot puts Claude on speaker QA for every line, then sends only the uncertain ones to people.
Auditable decisions
Every speaker assignment is stored with Claude's reasoning and confidence, so your QA team can see why a line went to a character.
Tunable confidence
Set the auto-apply threshold per project. Strict for a theatrical release, looser for a fast social cut.
Human review on the hard lines
Flagged lines route to your reviewers or ours. People spend their time on the 5% that needs judgment, not the 95% that doesn't.
API and volume
Run Claude-checked dubs across a full catalog through our API, with volume pricing and dedicated onboarding.
Speaker matching runs on Claude Sonnet 5.5 and Claude Haiku 4.5. Claude is a trademark of Anthropic, PBC. Monoglot is an independent company and is not affiliated with or endorsed by Anthropic.