Transcribe product

Transcribe is a local-first workbench for turning handwritten notebook pages into editable, portable text.

Is this for me?

Yes if you keep paper notebooks (or scans of them) and want searchable, editable text while keeping images and results on disk you control.

No if you need cloud OCR, audio transcription, or speaker diarization — those are out of scope. Transcribe does not depend on TranscriptX.

Promise

On your machine you can:

  1. Import JPEG/PNG/PDF pages into a notebook folder

  2. Transcribe them with a local Ollama vision model

  3. Review and correct text page by page

  4. Optionally analyse the text (Overview, Themes, Mood, Summaries, Detect, People & Places)

  5. Export Markdown, plain text, HTML/EPUB/PDF, and a portable notebook JSON file

  6. Back up and restore the full workspace as a local ZIP

Notebook identity, original scans, and human edits survive renames and re-OCR. Bulk import of many folders is supported. Contracts: notebook-corpus, source-asset, import-run, corpus-integrity.

Surfaces today

Surface

Role

Streamlit UI (port 8510)

Primary interactive workflow

CLI (transcribe / python -m transcribe)

Automation and integrity checks

Shared Python services

Single implementation for UI and CLI

Supported entrypoints: public_surfaces.md.

Product boundaries (v1)

In scope

  • Local Ollama vision models only (no cloud OCR providers)

  • Page-first domain (ordered pages, not timed speaker segments)

  • Human edits preserved separately from raw OCR attempts

  • Portable export without required absolute paths

  • Full-workspace backup / restore (transcribe.workspace-backup ZIP; replace-only onto current mounts)

  • Core notebook analysis modules and Analyse (optional local text Ollama for LLM modules)

  • Deepen-in-place: usability wave — trust, Analyse product UX, first-run operability (U2 open except Home/Diagnostics from GUI alignment), daily workbench (U3 done); OCR fail-fast, Analyse corpus-compare, Moments/chart jump → Reading, and Analyse/View split are shipped deepen-in-place (ROADMAP.md · usability_wave_plan.md)

Path to 1.0: package 0.8.8 (I0–I5 landed plus post-U3 product cuts and Docker-preferred install docs) → remaining U2 + I6 → cut 0.9.0 → 0.9-1 unfamiliar testing (dev/user_testing_0_9.md) → 1.0 freeze. Detail: ROADMAP.md Path to 0.9.0.

Out of scope for current core

  • Cloud OCR / hosted inference as a first-class provider

  • Audio transcription or speaker diarization

  • Shipping TranscriptX integration (future seam only — INTEGRATION_SEAM.md)

  • OpenCV-based preprocessing pipelines (optional Pillow profiles only; default is none). Visual declutter is a separate Pillow lane (scanner-bed, stark-white overscan, corner-wedge crop on import + explicit re-apply), not OCR preprocess.

  • Deferred analysis reinterpretations and ocr_quality — deferred on ROADMAP.md; prefer second-pass LLM OCR cleanup/verification for text quality

  • Autobiography / contextual imports (WhatsApp, photo libraries, Slices, reconstruction) — After 1.0 on ROADMAP.md; not current core

After 1.0 (planned)

1.0 remains this notebook/OCR/analysis workbench. Reach it via 0.9.0 (U2 + infra) then 0.9-1 unfamiliar testing (ROADMAP Path to 0.9.0). After that gate, Transcribe may grow into a local-first augmented autobiography workbench: handwritten notebooks stay the irreplaceable primary source; photographs, messages, transcripts, and mood records become evidence around them — never a replacement for the page.

Sequencing and architecture intent: ROADMAP.md After 1.0 (releases 1.1–2.0). This is not shipped behaviour and does not change v1 contracts. Foundation checklist (core freeze, ClaimStatus vocabulary, rebuildability, optional context trees) must be signed off at 1.0 before autobiography implementation.

Honesty

See known_limitations.md for model quality, PDF quirks, analysis capability caveats, and privacy. Shipped vs planned analysis: ROADMAP.md · analysis_wave1_plan.md.