oto.cx
transcription

mistral ai

transcription.

Start transcribing an audio recording (site visit, meeting, voice note)

ajouter à otoauth : —prefix : transcription_*

setup & usage

⚠ a Mistral key, training disabled

create an API key on the Mistral console, then paste it into oto at the organization level.

  • before a customer's first recording, disable the use of your data for training in the Mistral account settings: these are the voices of private individuals, often at home
  • by default each organization pays for its own audio minute at Mistral (EU hosting); an org without its own key may be granted a platform instance (that instance's key, language and vocabulary), and its own instance always takes precedence
    • ⚙ language and vocabulary

    two optional settings on the instance, non-secret:

  • language: the code of the spoken language (fr if empty). auto lets Mistral detect it
  • vocabulary: the business words to spell correctly (structures, materials, proper names), separated by commas or line breaks. Mistral only takes isolated words: an expression is split into its words, and words under 3 letters are discarded ("pompe à chaleur" → pompe, chaleur). 100 words at most; what is discarded is reported in the response

    • one instance = one key × one language × one vocabulary; a project attaches to the desired instance through a slot.

    ▸ a recording becomes a project page (asynchronous)
  • "transcribe the visit dropped on the project" → transcription_create(source={"kind":"project_file","project_id":…,"file_id":…}, _project=…) — the ids come from oto_project_files op=list
  • vocabulary (optional) supplements, for this recording, the instance's vocabulary (proper names, site materials); vocabulary_replace=true uses it alone. 100 words at most in total, the instance takes precedence
  • the file is read server-side, never carried through the conversation (100 MB at most, up to 3 h of audio; a project file remains limited to 25 MB at upload)
  • the call does not block: it immediately returns {job_id, status:"pending"} — a 30 min recording takes 20 s to 5 min to transcribe, as a background task
  • re-read transcription_status(job_id) until status:"done" (or "failed" with error): that is where the page "Transcription — <file> — <date>" comes back (id, title, link), one paragraph per speaker turn with the speaker and the moment at the head ("Locuteur 1 [03:12]"…), the word count, the duration and the speakers — never the text, which is then read like any project page
    • ▸ from a program (REST API, `oto_…` token)
  • the audio itself: curl -H "Authorization: Bearer $OTO_TOKEN" -F file=@reunion.ogg -F vocabulary="Voxtral, Otomata" https://mcp.oto.cx/api/me/projects/<id>/transcriptions/upload → 202 {job_id, status:"pending"}
  • a file oto already knows how to reach: POST /api/me/projects/<id>/transcriptions with {"source":{"kind":"project_file","file_id":…}} (or url, drive, gmail)
  • re-read GET /api/me/transcriptions/<job_id>: on done, in addition to the page, transcript returns the speaker turns as data — [{speaker, start, end, text}], in seconds
    • · what the connector corrects, and what it does not do
  • identically repeated segments are merged; the "parasite" speakers that diarization invents for a single sentence are attached to the neighboring speaker (at most 3 speakers are kept, each weighing at least 10 % of the recording)
  • nothing is transcribed without a call: you only pay for what is requested
  • the transcription draws no conclusion from the text: it is the project's procedure that says what to do with it (visit sheet, report…)

    • outils

    chargement…

    utilisation

    claude code — installe le plugin oto — claude plugin marketplace add otomata-tech/oto-plugin — mcp + skill configurés.
    claude.ai — branche mcp.oto.cx/mcp en connecteur custom (guide)
    cli — pipx install oto-cli puis oto transcription …
    ← tous les connecteurscatalogue mcp →