# Souffleur local transcription pilot

Souffleur installs all three pilot models and their CPU runtimes. Transcription stays on
this computer. The application does not download a model at startup and does not switch
to cloud transcription when a local model is missing or incompatible.

<!-- STT-PILOT-TABLE:START -->
| Selection | Mode and pilot status | Bundled model / runtime | CPU and RAM | Ryzen 7 7700X corrective measurement | Office-laptop evidence | Recommendation |
|---|---|---:|---|---|---|---|
| **Nemotron 3.5 ASR 0.6B Q8** | Retained-context live; **Recommended**; both short interjections retained in the focused live comparison | 741.5 MB / shared 6.6 MB NeMo-Speech closure | SSE4.2, AVX2, FMA, F16C and BMI2; RAM is reported, with no binding minimum yet | 15.43% isolated WER; RTF 0.1164 / 0.1160 | Not yet measured; #825 owns the representative office-laptop run | Automatic on every compatible pilot machine |
| **Qwen3-ASR 0.6B Q8** | Buffered final; **Pilot**; critical-term recall remains below the pilot gate | 1,151.3 MB / 7.3 MB audio.cpp runtime | Same instruction-set floor; RAM is reported, with no binding minimum yet | 13.19% isolated WER; RTF 0.1284 / 0.1285 | Not yet measured; #825 owns the representative office-laptop run | Explicit one-week pilot override |
| **Parakeet TDT 0.6B v3 Q8** | Final-only; **Pilot**; missed one of two short interjections | 714.0 MB / shared 6.6 MB NeMo-Speech closure | Same instruction-set floor; RAM is reported, with no binding minimum yet | 11.41% isolated WER; RTF 0.0602 / 0.0620 | Not yet measured; #825 owns the representative office-laptop run | Explicit one-week pilot override |
<!-- STT-PILOT-TABLE:END -->

The three models plus the exact runtime and notice closure occupy about **2.62 GB
decimal (2.44 GiB)** before the application and broker. Automatic is deliberately a
single tier: a compatible CPU receives Nemotron. A CPU missing any required instruction
is reported as incompatible. Physical-core and RAM values are evidence for work item
#825; they are not eligibility thresholds in this pilot.

These figures come from the #787 corrective release-binary benchmark and focused live
comparison. Aggregate WER alone did not choose Nemotron: the recommendation also uses
retained conversation context, clean speaker handoff and interruption behavior. The
pilot lasts one week. Work item #825 decides the binding fleet minimum before broader
rollout.

See `licenses/STT_THIRD_PARTY_NOTICES.md` in this directory for model and runtime
provenance, revisions, checksums and licence references.
