TranscrIA

TranscrIA is a self-hosted web portal that turns meetings into speaker-attributed transcripts and structured minutes, entirely on your own hardware. Upload a recording, capture from the microphone — or send a meeting bot into the call: Jitsi, Visio (LiveKit / La Suite numérique) and Zoom are supported, with per-participant audio tracks under each speaker's name and outbound-only connections that pass through corporate firewalls. Meetings can be scheduled from the UI and the bot joins on time. A guided wizard then walks through the pipeline: speech-to-text with a choice of engines (Whisper, Cohere, Voxtral, Kroko and more), speaker diarization, a shared lexicon for names and jargon, LLM-based correction and structured summaries — with a human validation step before anything becomes final. Attach the meeting's documents (PDF, DOCX, PPTX) and they feed both transcription and summary. Each job ends with a downloadable package: SRT subtitles (raw and corrected), timestamped segments, a structured summary, and Word minutes shaped by the meeting type. A built-in SRT editor with waveform, doubt indicators and a refinement chat lets you fix the transcript where it matters. Built for organizations, not just homelabs: enterprise identity (OIDC SSO, LDAP / Active Directory, trusted-proxy authentication, API tokens) with directory-driven group-to-role mapping, local accounts remaining the default. Deployment scales with you: single-box all-in-one, split web / scheduler roles sharing PostgreSQL, and remote GPU inference nodes for heavy STT — as Docker images (a fully offline one bakes all models in) or a native install with an express one-confirmation mode. Everything runs offline on local NVIDIA GPUs — 8 GB of VRAM is enough, and a CPU-only profile keeps transcription available without any GPU. No audio, transcript or summary ever leaves your infrastructure. Web UI in English and French. Apache-2.0.