Overview
keepcadence is an AI dubbing platform. It takes a source video or audio file, transcribes it, translates the transcript, synthesises speech in the target language using voice-cloned models, and re-times every sentence so the dub stays in cadence with the original speaker.
The four cadence mechanisms
Section titled “The four cadence mechanisms”- Length-hinted translation — the translation prompt includes the source sentence duration so the LLM targets a proportional word count.
- Per-sentence TTS duration target — each sentence carries an explicit duration hint fed to the TTS engine.
- Pitch-preserving time-stretch — output audio is stretched or compressed up to ±1.20× using a phase-vocoder that preserves perceived pitch.
- Forward drift into next silence — residual drift is absorbed by shifting the next silent gap rather than accumulating across sentences.
Architecture at a glance
Section titled “Architecture at a glance”| Layer | Technology |
|---|---|
| API | Hono on Bun (apps/api) |
| Reconciliation loop | Hono on Bun (apps/orchestrator) |
| Job queue | Postgres media_job table, FOR UPDATE SKIP LOCKED |
| ML workers | RunPod serverless GPU endpoints (Python) |
| Storage | Cloudflare R2 |
| Auth | Better-Auth |
This documentation
Section titled “This documentation”- Runbooks — step-by-step operational procedures: deployment, ASR, TTS, database backups, restore drills, and vendor integrations.
- Guides (this section) — conceptual overviews aimed at engineers onboarding to the codebase.
Full product documentation (concepts, API reference, user guides) is planned for after MVP launch.