ResearchTranscript

Formerly LocalTranscript, briefly TurnScript. Renamed in September 2026 because other transcription apps use similar names. Version counting restarts at 0.4.0, in line with the other B/IAS tools. ResearchTranscript starts fresh: it does not take over the settings of earlier versions, and on first launch you choose the library folder — an existing library can simply be picked there. Transcripts and dossiers written by the earlier versions are read as its own, and REFI-QDA exports keep their identifiers. Old links redirect.

A native macOS app for fully local audio transcription with speaker diarisation. Whisper runs on your machine, the diarisation runs on your machine, and nothing is ever uploaded. Built at B/IAS — Basel Institut für angewandte Stadtforschung as a Tauri app with the enrich UI kit, and it exports straight into enrich and into REFI-QDA for ATLAS.ti.

• Transcription: whisper.cpp (Metal), large-v3-turbo bundled.
• Diarisation: SpeakerKit (Argmax, MIT) running pyannote community-1 on Core ML — no Hugging Face token, no account, no network.
• Interface: German, English, French, Italian.
• No server: since 0.5.0 the Python logic runs inside the app process (embedded CPython 3.13 via PyO3); the interface calls it through Tauri commands. No port, no HTTP, no child interpreter. The loopback server (127.0.0.1:5628, LOCT on a phone keypad) exists only for browser development and the tests.

The screenshots below use an invented two-person interview about housing cooperatives, synthesised with macOS say. No real research material is shown anywhere in this repository. The file itself is in docs/demo/ — 2:02 min, two speakers — so the run can be reproduced without recording anything first.

Transcribe

Drop audio files (MP3, WAV, M4A, OGG, FLAC) onto the AI Transcript tab. Each file becomes a job with a live block counter, the text that is being recognised right now, and elapsed / estimated time.

Files run one after another by default — a second GPU job does not make the first one faster. Simultaneous runs in the settings raises that to at most four; the queue keeps the order in which files were dropped, and a waiting job can still be cancelled.

The bundled models are large-v3-turbo and medium. Any other whisper.cpp model — a Swiss German fine-tune, a quantised variant — goes into the library's Modelle folder and appears in the model list; see Your own models under Settings below. The app never downloads.

The Speakers option is an exact count, not a range: 1 means no separation and everything ends up as one speaker, 2–6 force exactly that many, Automatic lets the clustering decide. Separation sets how eagerly voices are split apart.

A library, not a pile of files

Every transcript is one folder holding the canonical transkript.json — speakers as entities, segments with stable ids — plus the audio, history/ snapshots written on every save, and the derived exports. Deleting moves to a library trash; nothing is destroyed.

Video

An MP4/MOV/M4V with H.264 or HEVC is accepted wherever audio is — in the batch list and as the audio of an import. The sound track is extracted to mp3 (the working copy for Whisper and the editor); the video file is stored unchanged in the entry and never transcoded. The only limit is 10 GB per file; other containers and codecs (webm/VP9, mkv, AVI, MPEG-2, AV1) are refused with a hint to export as H.264/MP4 externally, for instance with HandBrake or QuickTime. In the editor the picture sits under the speakers in the side panel, muted and without controls: the audio player leads, the picture follows — at 2× and during jumps it freezes and blurs until it has caught up. Faces are personal data; the privacy sheets say what that means for storage and export. The REFI-QDA export offers “with video” (VideoSource, video in the Media folder); the enrich dossier carries the sound only.

Edit

Click an entry to open the editor. Autosave, speaker rename / merge / reassign, an audio sample per speaker, split, join and delete segments, and a player that follows the text. Timecodes are always hh:mm:ss.

Keyboard, outside the text fields: J / K / L shuttle back, pause and forward the way Audition does — ← → and Space do the same; ↑ / ↓ walk along the segments and put the playhead on the start of each one. With ⌥ held, J / K / L also work while typing. Ctrl+L toggles the loop, Ctrl+X cycles the playback speed.

Metadata from Zotero

The side panel has a third tab, Metadata. If you keep your interviews in Zotero, link the transcript to its Zotero item: search by title, year, name or citekey (interviews come first), choose which people go into the transcript by role — the interviewee is not preselected, so a pseudonymised transcript does not get their name through Zotero — and link. Title, date, interviewer and citekey then appear on page 1 of the .enrich export and travel as the layer source/zotero.json. ResearchTranscript reads Zotero only on request, read-only (Zotero can stay open) and only after you enable it in Settings; nothing is sent to Zotero or the network.

Find and replace

The second tab in the side panel searches literally — no translation, no stemming, no fuzzy matching. It counts the occurrences, rings the segment it is currently on, and shows the match in its context. Replace changes one and moves on, Skip leaves one alone, Replace all does the rest.

Ignore hyphenation additionally finds words that a transcript broke across a line, so searching cooperative also finds coope- rative.

Export

• VTT — WebVTT with standard voice tags <v Name>.
• CSV, TXT — plain derived text.
• enrich dossier (.enrich) — one uncompressed file in Format 2 of the enrich dossier (FORMAT.md in the enrich repository): the transcript as the source with origin on every segment and speaker, the audio as mp3, the Zotero metadata if linked, and a manifest — written with enrich-core's own models — that inventories every file with its hash and carries the chained journal of who wrote what and when. No PDF: enrich typesets the reading copy itself on import (FORMAT.md §5), so the container stays small. Import it into enrich directly.
• REFI-QDA for ATLAS.ti (.qdpx.zip) — the transcript as a TextSource and as the Transcript of an AudioSource with one SyncPoint per utterance, speakers as codes. Verified against a project exported from ATLAS.ti 25.

Import — and the two steps for VTT

Import transcript in the Human Editor takes a transcript that was made elsewhere and turns it into a library entry.

A .enrich is one step. It is one file — a zip like .docx — that already carries its audio, so the app takes the transcript and the audio out of it and is done; the origin flags and the journal come along, and a container whose files no longer match their hashes is refused. The same works for a dossier folder as enrich keeps it (a package on macOS): pick it or drop it. Everything else in the dossier is dropped on purpose: after the first edit no analysis layer would line up with the text any more.

A .vtt or .csv is two steps, because those formats hold text and timecodes but no sound:

1. Pick the transcript. A file dialog opens for .vtt / .csv (and .enrich, including the .enrich.zip of earlier versions).
2. Pick the matching audio. A second dialog asks for the mp3 that belongs to it — Cancel if there is none. Without audio the entry still works; you simply edit text against timecodes and hear nothing.

If you skip the second step, the app looks for an audio file sitting next to the transcript with the same name (interview.vtt → interview.mp3) and takes that. So a matching pair in one folder imports in one step after all.

Handing a transcript to a colleague

The same .enrich does both jobs: it feeds enrich, and it hands a transcript to someone else who continues in their own Human Editor. Text and audio travel together in one file, the timecodes stay exact, and nothing is lost on the way out or back in.

Settings

Storage location, default model and language, interface language, simultaneous runs — and

Your own models. The library has a Modelle folder. Put any whisper.cpp model there (ggml-*.bin — large-v3-turbo, a quantised variant, a fine-tune converted with whisper.cpp's converter) and it appears in the model list, marked as your own; a file with the same name as a bundled model replaces it. The app never downloads models. A Hugging Face fine-tune is converted with scripts/hf-nach-ggml.py (whisper.cpp's converter, with the token table written by id — needed for models with a retrained tokenizer such as CrisperWhisper). Files are checked (ggml magic, finished copying) before they are offered, and Settings names what was ignored and why. Settings also show the full list of bundled components with their licences, the interchange formats with theirs, and the links to the sources. The same texts appear in the About ResearchTranscript dialog in the app menu.

Documentation

• This file — what the app is and does, architecture, development, building, licences.
• site/ — the project website: a static package with no cookies, no dependencies and no build step (fonts and screenshots in their own folders), carrying the user guide in German, English, French and Italian.
• CHANGELOG.md — version history. The GitHub release notes are the matching section from it; changes belong here first, not in the release form.
• In the app — Settings and the About dialog carry the storage location, formats, bundled tools and their licences, in four languages.

Architecture

backend/    Python ≥3.12 (uv): FastAPI, whisper-cli wrapper,
            diarisation, library, exports
            └─ src/localtranscript/enrich_export/  vendored enrich
               building blocks (drift-guard test) + Recursive fonts (OFL)
frontend/   React 18 + TS + Vite, Radix Themes, enrich kit
            (components/ui.tsx), i18n de/en/fr/it
            └─ src-tauri/   shell: spawns the backend, save dialogs
scripts/    bundle-resources.mjs (python runtime, whisper-cli, models,
            python/site-packages → src-tauri/resources), baue-lame.sh
            sign-resources.mjs, notarize-dmg.mjs

The shell is deliberately dumb: config, library and jobs belong to the backend. A backend already running in a terminal is used as it is and never touched; the app only shuts down what it started itself.

Development

Requirements: macOS on Apple Silicon, uv, Node 18+, Rust/cargo, brew install whisper-cpp (ffmpeg only for LT_MOTOR=kind), and ggml-large-v3-turbo.bin under models/ or ~/whisper-models/ (or LT_MODELS_DIR). An enrich checkout as a sibling (../enrich) — enrich-core is a path dependency of the .enrich export.

cd backend && uv sync && uv run pytest          # backend + tests
cd frontend && npm install
npm run dev                                      # browser dev (proxy :5628)
uv run uvicorn researchtranscript.main:app --port 5628    # in backend/ (dev server only)
node scripts/bundle-resources.mjs                # once: runtime + python/site-packages
node ../scripts/pyo3-config.mjs   # writes src-tauri/pyo3-config.txt for this machine (also done by bundle-resources.mjs)
PYO3_CONFIG_FILE=$PWD/src-tauri/pyo3-config.txt npx tauri dev   # app dev (Python in-process)

Building the bundle

node scripts/bundle-resources.mjs   # runtime/binaries/model + venv
node scripts/sign-resources.mjs     # 250 executables: Developer ID
cd frontend && npx tauri build      # .app + .dmg, shell sealed

The order is mandatory. Tauri signs the shell and the main binary only; the bundled executables and libraries (libpython, whisper-cli, libmp3lame, pydantic_core …) would otherwise keep the linker's ad-hoc signature — and Apple's notary service rejects those, after the 1.9 GB upload. sign-resources.mjs signs them with the Developer ID, hardened runtime, entitlements and a timestamp (55 s); the signature lives inside the Mach-O and survives being copied into the bundle.

The entitlements in frontend/src-tauri/entitlements.plist are not negotiable: without disable-library-validation the app does not start after signing, because Python loads .so files carrying a foreign signature.

Everything in one call (needs an app-specific password from appleid.apple.com):

export APPLE_ID=…  APPLE_PASSWORD=…  APPLE_TEAM_ID=CCRJ4A42D3
cd frontend && npm run release
spctl -a -vv /Applications/ResearchTranscript.app   # → Notarized Developer ID

npm run release chains the steps: stage the resources, sign the executables, tauri build --bundles app without the APPLE_* variables (signature only), notarize-app.mjs (submit with a two-hour timeout, staple), build-dmg.mjs (hdiutil + codesign) and finally notarize-dmg.mjs. Notarisation used to run inside tauri build; its internal wait expired on a slow Apple queue and took the whole build with it, so it lives in its own step now. Without that step Gatekeeper reports "Unnotarized Developer ID" when the downloaded image is opened, even though the app inside it is clean. If the ticket is already stapled the script does nothing and saves the 1.8 GB upload.

bundle-resources.mjs takes the python runtime, whisper-cli/dylibs and models from a v1 checkout sitting next to this one (../whisper-web/electron/resources) and builds the venv fresh. The dossier reader/writer comes from the public package enrich-core (MIT), pinned to a tag in backend/pyproject.toml — the same code the tests run against. Audio: since 0.6.0 the app decodes and probes media with Apple's AVFoundation and encodes MP3 with LAME (LGPL-2.0+), built from the official 4.0 tarball without decoder and frontend by scripts/baue-lame.sh and linked dynamically from Contents/Frameworks so it can be replaced. The tarball ships next to the library and is hosted at bias.city/researchtranscript/quellen/. ffmpeg is no longer bundled; LT_MOTOR=kind in a checkout still runs the old child processes (ffmpeg, argmax-cli) for comparison.

Interchange formats and their licences

REFI-QDA (.qdpx.zip). ResearchTranscript supports export to REFI-QDA. The specification is under the MIT licence, Copyright 2019 REFI-QDA (https://www.qdasoftware.org/). There is no official certification for REFI-QDA, and the MIT licence of the specification does not cover trademarks — no such claim is made here. The REFI schemas (XSD) are not bundled: ResearchTranscript writes to the specification and only references the schema URL, so the MIT attribution requirement does not apply.

enrich dossier (.enrich). The format is described in FORMAT.md — Format 2, which enrich and ResearchTranscript write today, and Format 1 for reading — and implemented by the reference package enrich-core, both under the MIT licence (B/IAS).

WebVTT (W3C) · CSV · TXT are open and unrestricted.

Citing

Pohl, Ben (2026). ResearchTranscript [software]. B/IAS – Basel Institut für angewandte Stadtforschung. https://doi.org/10.5281/zenodo.22831469

This DOI always resolves to the latest version; each release has its own version DOI on Zenodo (community BIAS.CITY). GitHub's "Cite this repository" reads CITATION.cff, and every documentation package the app exports contains the same entry as BibLaTeX for Zotero.

Authorship and AI assistance

UX/UI and software architecture were developed at B/IAS – Basel Institut für angewandte Stadtforschung. The code was written with the agentic coding capabilities of Anthropic's Claude Opus 5 and Claude Fable 5.1, directed, reviewed and tested at B/IAS. Responsibility for the software lies with B/IAS.

Licence

AGPL-3.0-or-later (BIAS.City), by choice. Up to 2.2.0 the licence was also forced by PyMuPDF (AGPL-3.0), which typeset the dossier PDF; since 2.3.0 the enrich export carries no PDF (enrich typesets the reading copy itself on import, FORMAT.md §5), so PyMuPDF and the Recursive fonts are no longer bundled; since 0.6.0 ffmpeg is gone as well. No bundled component requires the AGPL — the strictest is LAME (LGPL-2.0+, dynamically linked).

App Store: the same code is distributed through the Mac App Store under an additional permission (AGPL §7) in LICENSE-EXCEPTION; the source stays AGPL. Every bundled component is listed with its licence text in THIRD_PARTY_LICENSES.md (generated by scripts/gen-licenses.py, shipped inside the app, button "All licences" in the settings).

Texts and images are CC BY 4.0. The website in site/, the eight sheets in site/docs/, this README, the changelog and the backlog may be used and adapted, commercially too, as long as B/IAS is named and changes are marked — see LICENSE-docs. That is deliberate: the sheets exist to be pasted into an ethics application or a data management plan. The code stays AGPL-3.0-or-later, and the Recursive font on the website stays SIL OFL 1.1.

The network clause is satisfied before it applies: ResearchTranscript is not a network service — the logic runs inside the app process, and the only listener, the development server in main.py, binds to 127.0.0.1 and turns everything foreign away with 421 — there is no remote use in the sense of §13. The source is public anyway; the "Source code (GitHub)" button in the About dialog is the offer inside the app itself.

Bundled, among others: whisper.cpp (MIT), large-v3-turbo model (OpenAI, MIT), silero-vad (MIT), SpeakerKit (Argmax, MIT) with the pyannote community-1 and WeSpeaker ResNet34 models (CC BY 4.0, converted to Core ML and quantised by Argmax), enrich-core (B/IAS, MIT), CPython 3.13 (PSF), pydantic (MIT), React/Radix (MIT), Lucide (ISC), LAME (LGPL) — the complete list is in the settings.