Skip to content

Evidence pack — v0.0.4 residual (Group A)

Dated: 2026-07-31
Hardware profile: apple_silicon_metal (maintainer Apple Silicon Metal)
Aurum: 0.0.4

This document links the first retained, reproducible quality and performance artifacts after the v0.0.4 product-proof release. It is not a human multi-accent field study and not MOS.

Artifacts

Paths are in the repository root (not under docs/):

Kind Path
Public corpus evals/corpus.public-v1.json
Speech audio evals/audio/ (+ SHA256SUMS)
STT matrix evals/reports/stt/
TTS objective evals/reports/listening/tts-objective-apple_silicon_metal.json
TTS listening pilot evals/reports/listening/listening-round-001.json
Perf STT/TTS evals/reports/perf/

STT findings (synthetic English corpus)

Models: tiny-q5_1, base, small-q5_1 (local, warm Metal).

Model mean_wer (all fixtures) silence false positives
tiny-q5_1 ~0.49 3/3 silence-like fixtures
base ~0.47 3/3
small-q5_1 ~0.45 3/3

Observed behaviors:

  • Silence hallucination — all three models emit short hypotheses on silence (you) and non-speech tone (e.g. Thanks for watching! / BEEP!). Tracked via silence_false_positive.
  • Brand name — “Aurum” often becomes Oram/Orem; expected for rare tokens.
  • Digits — spoken letter/digit sequences collapse to compact forms (ABC123, 5551212); raw word WER looks high without digit-aware scoring.
  • Clean prose — fillers / multi-sentence fixtures often WER 0.0 at base/small.

Budgets (provisional)

Category Budget Action
Silence FP rate investigate if > 0 on empty references do not claim “no hallucination”
Speech-only mean WER on this corpus informational; not a ship gate yet refresh with licensed human speech before product claims
Experimental large-v3-q5_0 excluded from normal profiles keep experimental tier

TTS findings

  • Objective: Kitten and Kokoro samples non-empty, mono, non-clipped, finite PCM.
  • Listening pilot (round 001): single non-blinded listener; not MOS. Kokoro scored higher naturalness; Kitten remains default (lighter, MIT-safe default path). Homographs weak for both.

Performance (Apple Silicon Metal, warm)

End-to-end CLI wall time on tests/fixtures/sample.wav (~5.2 s audio):

Model p50 wall RTF p50
tiny-q5_1 ~0.59 s ~0.11
base ~0.80 s ~0.15
small-q5_1 ~0.88 s ~0.17

TTS short phrase (warm packs):

Model p50 wall RTF p50 (approx)
kitten-nano-int8 ~0.95 s ~0.27
kokoro-82m-int8 ~1.84 s ~0.89

How to reproduce

cargo build -p aurum-stt --release
./scripts/generate_eval_speech.sh   # macOS say + ffmpeg
./scripts/generate_eval_audio.py
AURUM_BIN=target/release/aurum python3 scripts/run_stt_eval_matrix.py
# TTS + perf: re-run maintainer scripts or extend run_perf_report.sh

Honest limits

  • Corpus speech is synthetic (macOS say / lavfi), not licensed human speech.
  • Multi-accent coverage is synthetic voice variation, not real speakers.
  • Listening is a pilot, not a formal study.
  • One hardware profile only; Linux/Windows retained reports still needed.