{% extends "base.html" %} {% block title %}ArtificeTranscribe — About{% endblock %} {% block content %}
Speech-to-text · diarization · oral history

About ArtificeTranscribe

ArtificeTranscribe turns an audio or video recording into an editable, speaker-labelled transcript — the workflow an oral history project needs, running entirely on your own machine. Drop in a file, and Whisper (via WhisperX) transcribes it while pyannote separates who spoke when; the result is a segment-by-segment transcript you can correct, name, and export in the formats an archive expects.

What a recording goes through

  1. Transcribe — a Whisper model (tiny through large-v3, chosen in Settings · Models) turns speech into timestamped text.
  2. Diarize — pyannote segments the audio by speaker and aligns each word to a speaker turn, honouring a min/max speaker count if you know it in advance.
  3. Identify — name each detected speaker, or match against speakers already enrolled from an earlier recording so recurring interviewees don't have to be renamed every session.
  4. Edit & export — correct segments in place, with diff and edit history behind every change, then export as JSON, SRT, VTT, TXT, Markdown, PDF, OHMS or TEI.

Vocabulary and known speakers

The Dictionary tab holds a persistent list of proper nouns, jargon and place names that gets merged into every transcription as hotwords, so project-specific terms don't have to be corrected by hand each time. A short enrollment clip teaches the app a speaker's voice once; later recordings can then be matched against that voice rather than re-identified from scratch.

Interview metadata and the AI assistant

Each transcript carries its own interviewee, interviewer, date, location, project and collection fields, plus an access note, alongside the transcript itself. An optional BYOM assistant — any OpenAI-compatible endpoint, local or cloud — can summarise or clean up a finished transcript on request; it never touches the transcription or diarization models themselves.

Design language

This interface implements The New Masses design system: an offline-first, print-inflected editorial aesthetic with warm paper surfaces, a muted violet accent, and fluid serif typography. All design tokens are served from the shared packages/shared-ui/shared_ui/assets/tokens.css; never hardcode colours or spacing.

Run it

From the project root: 

pip install -e ".[web]"
artifice-transcribe
# → http://localhost:8000

ArtificeTranscribe · an Artifice platform · FastAPI + Jinja2 · vanilla JS.

{% endblock %}