TranscrIA

August 15, 2026 · View on GitHub

Reference documentation for operators, integrators, and contributors. The product itself is bilingual French / English (interface, deliverables, installer, doctor — pick the language at install time or from the navbar). The project README is available in both languages (English · français) and covers the full install + Docker quickstart. These deeper reference documents are written in French; this index is the English entry point (English summaries below).

New here? The single-page walkthrough QUICKSTART.en.md (français : QUICKSTART.md) takes you from zero to your first meeting minutes; then INSTALL.en.md or DOCKER.en.md — the exhaustive French references being INSTALL.md and DOCKER.md. Non-technical readers (business owners, project managers, decision makers) have a dedicated overview: PRESENTATION.md (French) · PRESENTATION.en.md (English) — use cases, benefits, example results and the user journey.

Deployment and operations

DocumentWhat it covers
TESTERS.mdIn English — testing TranscrIA: what to expect (disk, first startup, models, GPU floor), the 15-minute smoke test, topologies we need tested, diagnostics and the report template
QUICKSTART.md · QUICKSTART.en.mdFrom zero to your first minutes on one page — host prerequisites, one Docker command or the express native install, first login, first job
QUICKSTART_WINDOWS.md · QUICKSTART_WINDOWS.en.mdWindows 11 — guided install script + WSL2/Docker Desktop path (GPU passthrough, .wslconfig, disk/VHDX pitfalls); verified against official docs, real-machine validation in progress
INSTALL.md · INSTALL.en.mdHost installation (install.sh), hardware and CUDA detection, models, systemd service, distributed roles, troubleshooting — English version condensed, French exhaustive
DOCKER.md · DOCKER.en.mdContainerized deployment: turnkey quickstart, slim vs. bundled images, Compose, GPU access, variables, rollback
UPGRADE.mdUpgrade and rollback procedure, obsolete configuration keys, database migrations
RELEASE.mdMaintainers — how a version gets published: gates, real E2E, clean-distro install gate, Docker rule, documentation review, tag, images, release notes, and the incidents each step exists for
SERVICE_RESSOURCES_GPU.mdSplit topology (web frontend + GPU resource node): remote inference, VRAM autonomy, admission and degraded modes
STOCKAGE_PARTAGE_JOBS.mdPostgreSQL-backed job file store for deployments without a shared filesystem

Reference

DocumentWhat it covers
TECHNICAL.mdArchitecture, pipeline, modules, GPU orchestration, timing model, database
API_REFERENCE.mdGenerated HTTP API reference (all routes, auth, ⭐ scriptable contract) — regenerate with scripts/generate_api_reference.py, guarded in CI
CONFIG_REFERENCE.mdComplete config.yaml reference — hand-written, but key coverage is guarded in CI (release_check.py: every schema key must have its line)
DATA_MODEL.mdDatabase schema, job states, per-job files
LLM_TIERS.mdIn English — the benchmarked arbitration-LLM VRAM tiers (12 → 64 GB): validated model, quant and context per tier, backend recommendation
LLM_BACKENDS.mdArbitration LLM backends (Ollama / llama.cpp / vLLM) and hardware-driven selection
I18N_MULTILANGUE.mdBilingual FR/EN architecture (interface, deliverables, installer, doctor) and how to add a language

Features

DocumentWhat it covers
PROFILS_TRAITEMENT_WORKFLOW.mdThe six processing profiles and the human-in-the-loop wizard
TYPES_REUNION_PERSONNALISES.mdCustom meeting types: catalog, detection, extracted fields, DOCX theming
EDITEUR_SRT_INTEGRE.mdBuilt-in SRT editor: versioned transcript correction
PIPELINE_REPRISE.mdResumable pipeline: phase checkpoints and provenance fingerprints

Meeting connectors

TranscrIA can fetch a meeting on its own — without any inbound port. Four platforms are validated in real conditions; two more ship prepared but unvalidated.

DocumentWhat it covers
TEMPS_REEL_REUNIONS.mdMaster plan: the three ingestion routes (post-meeting artefacts, live media, STT client) and the per-platform study
BOT_REUNION.mdThe meeting bot: lifecycle, states, capture, displayed identity, guards
VISIO_ZOOM_RUNNER.mdVisio (native LiveKit client) and Zoom (official SDK) on the full journey, with automated gate benches
ZOOM_GRATUIT_ADMIN.mdZoom admin guide: creating the Meeting SDK app on a free or business account, diagnostics
MEET_TEAMS_ADMIN.mdGoogle Meet and Teams admin guide: permissions, the SILENT failure modes, recommended order
RUNNER_DISTANT_KIT.mdInstalling a meeting runner on another machine (script generated from the admin UI)

Google Meet needs no bot at all — it collects the recording afterwards through a Pub/Sub pull. Zoom RTMS and Microsoft Teams are the only two routes that require a public HTTPS entry point; they are shipped implemented, not validated.

Security and compliance

DocumentWhat it covers
SECURITY_MODEL.mdAuthentication, RBAC, rate limiting, security headers
AUDIT_DPO.mdAudit trail, retention and purge, data-protection posture

Speech-to-text and audio

DocumentWhat it covers
STT_ADAPTATIF_ET_HYBRIDE.mdAdaptive and hybrid STT: quality-driven backend selection
EXTERNAL_STT_RUNTIMES.mdServed STT runtimes — audio.cpp (qwen3asr) and parakeet.cpp (nemotron) as first-class engines: pinned installer builds, on-demand start before jobs, per-engine health, native fallback

Benchmarking and validation

DocumentWhat it covers
BENCHMARKING.mdHow to run a reproducible bench: the bench_audio.py matrix runner and its four analyzers (metrics, LLM quality, WER vs reference, concurrency)
STT_BENCHMARK_REAL_MEETINGS.mdIn English — published STT benchmark on real French meetings vs a professional human transcript: all engines and external runtimes, traps, failure modes
BENCH_LLM_PALIERS.mdPer-VRAM-tier model benchmarks (the source for tier selection)
STT_CORPUS.mdHow the reference corpus is built (selection, annotation, licensing)

Concurrency, scale, and distributed inference

DocumentWhat it covers
MIGRATION_API_SERVEUR_GPU.mdRemote GPU resource-node HTTP API semantics (inference_service) — route list lives in API_REFERENCE.md
PLAN_TEST_CHARGE.mdLoad-test procedure and campaigns (mandatory net for GPU-concurrency changes)
PLAN_TEST_SPLIT_VLLM.mdSplit-topology (vLLM) validation procedure

Engineering plans

DocumentWhat it covers
REFACTORING_QUALITE.mdCode-quality master plan: measured state (god modules, import graph, hotspots), target layering, refactoring waves and permanent guardrails
PISTES_AMELIORATION.mdPost-0.3.7 improvement analysis: measured time breakdown, engine choices, CPU-fallback trade-offs, UX and operations gaps, suggested roadmap
GESTION_IDENTITE.mdEnterprise identity (shipped in 0.3.9): pluggable backends — OIDC, direct LDAP/AD, trusted proxy headers, API tokens — group→role mapping, JIT provisioning, break-glass, test matrix

History

Superseded planning documents, benchmarks, and analyses are kept under archive/ for provenance. They do not describe current behaviour; use the documents above instead. Recently archived (2026-07-16): the Phase B concurrency plan (CONCURRENCE_ET_CHARGE_PHASE_B.md, delivered — decisions absorbed into AGENTS.md), the Parakeet integration scoping (PARAKEET_STT_INTEGRATION.md, superseded by EXTERNAL_STT_RUNTIMES.md), the VAD study (VAD_OR_NOT.md, decision record) and the LLM tier validation records (LLM_PROFILS_VALIDATION.md — the living protocol is BENCH_LLM_PALIERS.md).