Skip to content

PROGRESS.md

Latest commit

 

History

History
48 lines (34 loc) · 3.8 KB

PROGRESS.md

File metadata and controls

48 lines (34 loc) · 3.8 KB
Raw

Subber development status

Last updated: 1 September 2026.

Subber’s local transcription/export core and the first real AI-assisted source review are working. Development is currently focused on making the standalone editing flow robust before commercial services or the DaVinci Resolve plugin are designed.

Implemented and manually exercised

  • Windows/macOS Tauri project workflow, persistent queues and per-video export folders.
  • FFmpeg audio extraction and faster-whisper transcription with automatic language detection.
  • Local NLLB contextual translation, glossary protection and formatted TXT/SRT/JSON output.
  • Editable transcript, reversible segment inclusion/exclusion, local ASR diagnostics and conservative VAD.
  • Word-aligned subtitle timing for new transcriptions: internal pauses split, short speech padding and non-overlapping SRT cues. This hardening is implemented and awaiting manual verification.
  • Styled caption preview and burned-in video export up to 4K.
  • Optional OpenAI BYOK source audit: project context, versioned prompt, structured output, inline suggestions and persistent non-destructive reports.
  • AI report reconciliation after structural edits and explicit batch action Apply all; saving does not trigger another API call.
  • Retranscription preserves previous AI report sidecars but marks their segment anchors stale before replacing the transcript; no automatic API rerun.
  • DaVinci bridge with explicit capability diagnostics: direct Media Pool import through the Resolve Studio external scripting API, in-place refresh of an already imported canonical SRT, duplicate-video prevention, and guided manual fallback for Resolve Free.
  • Canonical save/import contract: source and translated SRT paths are distinct, unsaved editor drafts cannot be imported, and every save overwrites the single current deliverable instead of creating revisions.
  • Source corrections automatically refresh only the affected local translations that were not manually edited before the final SRT is published; no OpenAI call is made.
  • Clean project layout: only video-sub.json and exports/ are user-facing; WAV, ASR/translation sidecars and AI reports live under .subber/.

Privacy boundary

The media pipeline is local. Video and audio are never sent to OpenAI. The optional AI source audit runs only after explicit consent and sends text, timestamps, ASR confidence, project context, glossary and preset using the user’s own API key with provider storage disabled.

Remaining roadmap

  1. Non-destructive AI suggestions for editorial segment selection.
  2. AI refinement of translations for retained segments.
  3. Subtitle QA and verifiable editorial presets.
  4. Quality, cost, privacy, Windows and macOS validation.
  5. Subber Pro/Studio architecture: backend, accounts and AI quotas.
  6. Standalone stabilization and a DaVinci Resolve plugin that reuses the same core.

The standalone core is advanced; commercial infrastructure and the Resolve plugin are still substantial future phases.

Current manual verification

The DaVinci bridge changes have not been built or run by the coding agent. Verify both supported paths:

  1. Resolve Free selects the SRT in Explorer/Finder and explains that manual drag-and-drop is required.
  2. Resolve Studio with External scripting using: Local and an open project imports the SRT into the current Media Pool folder.
  3. After editing and saving again, a second import refreshes that same Media Pool item instead of creating a duplicate; verify that existing timeline instances reflect Resolve’s refreshed media.

See docs/davinci.md or docs/davinci.it.md.

For timing verification, retranscribe a video after this change: old sidecars do not contain word timestamps. Compare speech onset, speech end, internal pauses and consecutive cues in Subber and Resolve.