Skip to content

Models

aurum models

Lists the local whisper.cpp catalogue, cache status, and support tier. Every trusted entry has a reviewed SHA-256 pin and exact size (JOE-1645).

Use case Model Notes
First-run trial tiny-q5_1 ~32 MB, fast
Draft / default speed base (default) or base.en-q5_1 Balanced size; default prioritizes speed
English lecture quality small.en or large-v3-turbo Best observed quality/speed in internal dogfood
Higher EN accuracy medium.en Slower; strong on long EN speech

Default remains base unless you opt into a larger model.

Quality evidence (JOE-2216 observatory)

The versioned STT quality observatory is the product evidence programme:

  • Methodology: docs/operations/stt-observatory.md
  • Core corpus + budgets: evals/observatory/
  • Evidence version: 0.0.22-observatory-v1 (PROFILE_EVIDENCE_VERSION)
  • Retained reports: evals/reports/stt/

The redistributable core exercises schema, silence controls, and budget comparison in CI. The production pack (≥60 minutes licensed speech, multi speaker/accent/noise/long-form) is prepared on operator machines from open licensed sources — never private Plaud material. Profile mappings cite the reviewed evidence version; explicit --model always wins. The global default remains base until a separate reviewed decision changes it.

Corpus evidence is not a claim of universal field quality. Scenario-level errors matter more than a single aggregate WER.

Historical single-clip dogfood (pre-observatory) still informs guidance only:

  • English-only (.en) variants often beat multilingual twins on EN lectures.
  • small.en and large-v3-turbo were sweet spots; full large-v3 was not automatically better.
  • large-v3-q5_0 is experimental / not recommended after severe repetition degeneration on that clip (see below).

Support tiers

Tier Meaning
supported Normal catalogue entry with full SHA-256 authentication
experimental Available but not recommended for production quality

large-v3-q5_0 (experimental)

Marked experimental after a pathological repetition loop in dogfood (~21% WER vs reference with a short phrase repeating dozens of times). Prefer large-v3-turbo or large-v3-turbo-q5_0 instead. Aurum’s normalization report API can detect extreme n-gram loops as degeneration_repetition; ordinary CLI/JSON transcript output does not currently surface that event unless hosts consume NormalizationReport directly.

Catalogue (aliases)

Name Approx size Notes
tiny-q5_1 ~32 MB Best first-run trial
base-q5_1 ~60 MB Good default quantized
base ~142 MB Default full-precision
small / small.en ~444 MB Higher accuracy
medium / medium.en ~1.4 GB Large download
turbo / large-v3-turbo ~1.5 GB Fast large
large-v3-q5_0 ~1.0 GB Experimental — not recommended

Aliases: large → large-v3, turbo → large-v3-turbo.

Trust and authenticity

  • Every trusted catalogue file has immutable SHA-256 + exact size pins.
  • Downloads verify-before-publish; wrong digest/size never becomes the final path.
  • A local .sha256 sidecar never authenticates an unpinned artifact.
  • Identity is the digest, not a mutable upstream branch tip.

Cache location

Platform Path
macOS ~/Library/Caches/aurum/models/
Linux ~/.cache/aurum/models/
Windows %LOCALAPPDATA%\aurum\cache\models\

First use downloads from Hugging Face (ggerganov/whisper.cpp). Cross-process locks prevent double downloads.

Offline / Local Only

let p = LocalWhisperProvider::new(cache).with_local_only(true);
// Fails with ModelNotCached if the file is missing — no network.