TranscrIA
Meeting minutes you can actually trust.
TranscrIA turns meeting recordings into speaker-attributed transcripts, LLM-corrected text and structured Word minutes — entirely on your hardware. Nothing leaves your network, and nothing becomes final until a human says so.
Automatic where machines are good.
Human where it matters.
Every fully-automatic minutes tool eventually attributes a decision to the wrong person. TranscrIA is built around two explicit checkpoints where you keep the last word.
Drop a recording — or send the bot
Upload any audio, or let the meeting bot join Jitsi, self-hosted Visio or Zoom for you. No inbound port required.
Acoustic preflight
Signal metrics, DNSMOS perceptual scores and a difficulty timeline — the verdict before you burn GPU hours.
Transcription + diarization
Pick your engine — all benchmarked on real meetings, including a CPU-only one that needs zero VRAM.
You validate the speakers
A guided wizard with waveform playback: name who's who, fix the machine's guesses, in minutes.
You approve the corrections
A local LLM proposes fixes and summaries; every change is traceable and nothing applies without you.
Word minutes + audit trail
Structured DOCX by meeting type, refine-by-chat if needed, GDPR audit trail included.
A portal, not a script.
Audio preflight verdict
Know if a recording will transcribe well before spending a single GPU minute. Try this exact step in your browser.
Meeting bots, no inbound ports
Jitsi, self-hosted Visio (LiveKit) and Zoom — validated in real conditions. The bot joins, records, and the pipeline takes over.
Guided review wizard
Built-in SRT editor with waveform, versioned corrections, and a review flow that matches how secretaries actually work.
Processing profiles
From “fast draft” to “thorough with final re-read” — the profile you pick at upload drives the whole pipeline.
From 8 GB to multi-GPU
One VRAM-aware queue, benchmarked LLM tiers from 12 to 64 GB, and a split topology when one machine isn't enough.
Enterprise-ready identity
OIDC, LDAP/AD, trusted-proxy headers, role mapping — plus a GDPR audit trail and retention rules.
We benchmarked 9 STT engines on real meetings.
Against a professional human transcript — hallucination counts and who-said-what errors included. We published the parts that don't flatter anyone, ours included.
Before you ask.
- French-first, English complete. German, Spanish and Italian shipped in beta — native speakers welcome.
- You need an NVIDIA GPU (8 GB VRAM is enough since 0.4.2; RTX 50xx supported). A CPU-only speech-to-text engine exists, but the correction LLM wants a GPU.
- Windows 11 runs through WSL2. The guide is written; we're looking for the first real-machine run.
- Google Meet and Teams connectors ship prepared, not validated. We say so out loud instead of putting logos on this page.