Skip to content

TranscrIA 0.3.5 — new engines & a smarter editor

Choose a tag to compare

View all tags
@Martossien Martossien released this 12 Jul 14:06
v0.3.5
c01402a

New engines & a smarter editor

Two new STT backends join the catalog — both qualified on our published benchmark of real French meeting recordings (docs/STT_BENCHMARK_REAL_MEETINGS.md, now covering 15 engine/runtime configurations) — and the transcript editor learns to keep the synthesis honest after human corrections.

New

  • Kroko-ASR backend — the no-GPU option (models.stt_backend: kroko): per-language streaming Zipformer models (~155 MB each, 10 languages incl. FR/EN) on sherpa-onnx, CPU only — zero VRAM reserved. On our real-meeting corpus it matched our best GPU engines (WER 0.43 vs 0.427 for the leader) at ~30× real time on 8 CPU threads, with punctuated, cased output. Community weights are CC-BY-SA: downloadable from the Models page (not baked into images). A language-dedicated small model beating language-generalist giants is one of the benchmark's main lessons.
  • MOSS-Transcribe-Diarize backend (models.stt_backend: moss, Apache-2.0, 0.9B): transcription + speaker labels + fine timestamps in a single pass — best text WER of the whole benchmark (0.41), and the only unified audio LLM we tested that survives real 5-minute meeting windows without looping. Requires Transformers 5.x → runs in an isolated subprocess worker (transcria.installer.cli moss-site, project venv untouched). Its one measured vice — silent omission (a 22 s skip we caught by reading outputs, invisible in WER tables) — ships with its own guard: inter-segment gaps are flagged in metadata, never auto-corrected.
  • Editor: quick DOCX or LLM-updated synthesis. After saving SRT corrections (sentences, speakers), the editor now offers a choice: download the DOCX immediately (verbatim and speaker stats already follow the SRT), or run an LLM pass that updates the summary/decisions only where the corrected transcript contradicts them — proposed, never automatic, versioned and restorable from the refine chat. Verified end-to-end on the real UI: the pass propagated an edited figure into the summary and even caught a pre-existing summary/verbatim mismatch on its own, with justification.
  • :bundled Docker image ships MOSS baked in (weights + the isolated Transformers 5 site at /opt/transcria-moss-site, symlinked at startup) — switching to models.stt_backend: moss works fully offline.
  • External C++ runtimes documented (docs/EXTERNAL_STT_RUNTIMES.md): audio.cpp and parakeet.cpp plug into the existing OpenAI-compatible remote STT endpoint with configuration only, zero code — both qualified on the benchmark with very good surprises.
  • Editor saves now rebuild the export ZIP (it no longer serves a DOCX predating your corrections); bench matrix extended to 42 combos (7 backends).

No database migration. New Python dependency sherpa-onnx (self-contained wheel, installed by pip install -r requirements.txt / install.sh / Docker images). Kroko and MOSS models are opt-in downloads (Models page). Full details: CHANGELOG · upgrade notes: docs/UPGRADE.md.

🤖 Generated with Claude Code

Assets 2
Loading