Will your meeting audio transcribe well?

Every transcription tool lets you discover after the GPU hours that your recording was unusable. This Space runs the acoustic preflight step of TranscrIA, a self-hosted meeting transcription portal, and gives you the verdict first — entirely in your browser:

• deterministic signal metrics — SNR, clipping, bandwidth, silence ratio;
• DNSMOS P.835 perceptual scores (SIG/BAK/OVRL), running as WebAssembly via onnxruntime-web;
• a per-window difficulty timeline showing where the recording will hurt;
• and since the verdict deserves a proof: speech-to-text in your browser — Kroko-ASR community models (English and French, CC-BY-SA) running on one CPU core via sherpa-onnx compiled to WebAssembly. These are the exact same model files TranscrIA ships as its zero-VRAM kroko backend — the ones measured in its public benchmark on real meetings, where they land near the big GPU models. Around 7× real time on a desktop CPU, in the browser.

Your audio never leaves your browser. No upload, no server, no API — this is a static page; decoding (WebAudio), metrics (JavaScript) and the neural model (WASM) all run client-side. Fitting, for a project whose whole point is data sovereignty.

Faithfulness to the real pipeline

js/preflight.js and js/dnsmos.js are line-by-line ports of TranscrIA's transcria/audio/preflight.py and dnsmos_scorer.py (same formulas, same thresholds, same flag names), validated by automated comparison against the Python outputs on the bundled examples (SNR within 0.1 dB, DNSMOS within 0.01 MOS). Browser-edition differences, on purpose:

• audio is decoded/resampled to 16 kHz mono by WebAudio (the Python ffmpeg fallback path does exactly the same);
• SQUIM (STOI/PESQ/SI-SDR, a PyTorch model) cannot run in a browser — the difficulty timeline is carried by per-window DNSMOS, SNR and clipping instead. The full version in TranscrIA adds SQUIM on top.

The transcription demo follows the same discipline: js/kroko.js is a port of the container parser from TranscrIA's transcria/stt/kroko_transcriber.py, the recognizer uses the same sherpa-onnx parameters as the validated Python backend, and the browser output was validated against the native backend on the bundled examples (bag-of-words F1 ≥ 0.99). The WASM runtime is our own build of sherpa-onnx v1.13.5 without the usual preloaded model (see build/build-sherpa-wasm.sh) — models download on demand (~156 MB per language, cached by your browser for next time), and the demo transcribes the first 2 minutes. Raw single-pass output, on purpose: no punctuation model, no speakers, no LLM correction — that is exactly what TranscrIA's pipeline adds on top.

What happens after the verdict?

In TranscrIA, this preflight gates a full pipeline on your own hardware: speech-to-text, speaker diarization, human-validated LLM correction, structured summaries and Word minutes — self-hosted, GDPR audit trail included.

Licenses and attributions

• Code: Apache-2.0 (see LICENSE).
• DNSMOS model dnsmos_sig_bak_ovr.onnx (Microsoft DNS-Challenge): CC-BY-4.0 — see models/DNSMOS_MODEL_LICENSE.txt.
• Kroko-ASR community models (Banafo): CC-BY-SA, redistributed unmodified — see models/kroko/KROKO_MODEL_LICENSE.txt (sha256 matching upstream).
• sherpa-onnx: Apache-2.0 — WASM runtime built by build/build-sherpa-wasm.sh (v1.13.5 pinned), wrapper vendor/sherpa/sherpa-onnx-asr.js unmodified.
• onnxruntime-web: MIT (bundled in vendor/ort/).
• Examples: clean_speech.mp3 — excerpt from a LibriVox recording (public domain); archival_1939.mp3 — Lou Gehrig, "Farewell to Baseball" (1939, public domain); degraded_speech.mp3 — the LibriVox excerpt artificially degraded (noise, band-limiting, clipping).