openquack

OpenQuack benchmarks

Living document. We add a row to the host matrix every time someone runs the bench and submits the result. Numbers below are real, reproducible from bench/out/<host-tag>/, and supersede earlier preliminary runs.

TL;DR (M4 / 16 GB)

WhisperKit medium is the recommended v2 default for English on 16 GB+ Macs. 2.6 % WER on LibriSpeech, 1.3 % on multi-voice English TTS, 6.3 % on noise-augmented speech. RTF 0.22–0.31× (≥3× realtime). Peak RSS 197 MB. Cold start 27 s (cache miss only — once per Mac).

For 8 GB Macs, whisperkit small is the practical choice (165 MB RSS, 4.1 % LibriSpeech WER, but degrades to 11 % on noise).

Non-English auto-detect works. WhisperKit medium transcribes the multilingual bucket at 16.7 % WER / 3.7 % CER on auto-detect — on par with Lightning, and a non-issue for non-English users. (Builds before alpha.17 shipped a config bug that left detection off and translated non-English audio to English; fixed in SPEC-021.)

Hosts

Host tag	Chip	GPU	Memory	macOS	Date
M4-16GB	Apple M4	8-core	16 GB	15.6 (24G84)	2026-04-26

Methodology

• Engines: WhisperKit (argmaxinc/argmax-oss-swift 0.18+, primary) and Lightning (lightning-whisper-mlx 0.0.10 via long-running Python subprocess, comparison-only).
• Corpus: 177 WAV clips, 16 kHz mono PCM, across five buckets:
  • librispeech/ — 20 real human read speech clips from LibriSpeech dev-clean (CC BY 4.0, openslr.org/12). The closest thing in this set to real-world dictation accuracy.
  • voices/ — 20 multi-accent English TTS (Samantha US, Daniel UK, Karen AU, Fred US-robotic, 5 sentences each).
  • noisy/ — 120 noise-augmented clips (the 20 voices × white + pink noise × 5 / 10 / 20 dB SNR).
  • multilingual/ — 12 native-language clips (zh, ja, ko, es, fr, de × 2).
  • short/ — 5 single-voice TTS sanity-check clips.
• Models: tiny, base, small, medium, distil-large-v3. (large-v3-turbo is unavailable in either engine under that name; tracked as an open question.)
• Metrics:
  • WER — Levenshtein over normalised whitespace tokens.
  • CER — same on characters. More meaningful for CJK; we report it everywhere for comparison.
  • RTF — wall_seconds ÷ audio_seconds.
  • Cold start — wall-clock from engine init to first transcribe ready (includes cache-miss download).
  • Peak RSS — sampled at 100 ms via mach_task_basic_info.
• Auto-detect: no language is hinted to the engine — the app’s default path, and what the multilingual numbers measure.

Per-bucket results — M4-16GB (2026-04-26)

LibriSpeech (real human speech, English read) — 20 clips

Engine	Model	WER	CER	RTF	Wall (avg)
whisperkit	medium	2.6 %	1.3 %	0.22×	1.98 s
lightning	medium	2.8 %	1.3 %	0.34×	2.47 s
whisperkit	distil-large-v3	3.2 %	1.3 %	0.12×	0.89 s
lightning	distil-large-v3	3.2 %	1.3 %	0.48×	3.27 s
whisperkit	small	4.1 %	1.7 %	0.05×	0.50 s
lightning	small	4.6 %	2.3 %	0.10×	0.74 s
lightning	base	5.6 %	2.8 %	0.04×	0.29 s
whisperkit	base	5.9 %	2.3 %	0.02×	0.20 s
whisperkit	tiny	7.2 %	3.0 %	0.01×	0.13 s
lightning	tiny	8.7 %	4.4 %	0.02×	0.18 s

Read: 2.6 % WER on real human read speech is solid — close to what dictation feels like in practice. medium outperforms distil-large-v3 here; small is acceptable but visibly less accurate.

Voices (multi-accent English TTS) — 20 clips

Engine	Model	WER	CER	RTF
whisperkit	medium	1.3 %	0.3 %	0.31×
lightning	medium	2.0 %	0.4 %	0.63×
whisperkit	small	2.7 %	1.2 %	0.09×
lightning	small	3.1 %	1.6 %	0.19×
whisperkit	distil-large-v3	5.3 %	0.7 %	0.23×
lightning	distil-large-v3	5.3 %	0.7 %	0.97×
whisperkit	base	5.6 %	2.1 %	0.04×
lightning	base	9.4 %	3.7 %	0.06×
lightning	tiny	12.9 %	5.1 %	0.04×
whisperkit	tiny	15.0 %	5.8 %	0.02×

Noisy (white/pink × 5/10/20 dB SNR over voices) — 120 clips

Engine	Model	WER	CER	RTF
whisperkit	medium	6.3 %	2.7 %	0.31×
lightning	medium	6.3 %	2.9 %	0.63×
whisperkit	distil-large-v3	9.4 %	3.5 %	0.23×
lightning	distil-large-v3	10.0 %	3.7 %	0.98×
whisperkit	small	11.4 %	5.5 %	0.10×
lightning	small	11.8 %	5.7 %	0.19×
lightning	base	22.8 %	12.5 %	0.06×
whisperkit	base	24.0 %	13.2 %	0.04×
whisperkit	tiny	27.3 %	15.8 %	0.03×
lightning	tiny	28.7 %	16.5 %	0.04×

Read: Once you add real-world noise, medium pulls clearly ahead of small. tiny and base collapse — they’re not viable in any non-pristine environment.

Multilingual (auto-detect, 12 clips: zh/ja/ko/es/fr/de × 2)

Engine	Model	WER	CER
whisperkit	medium	16.7 %	3.7 %
lightning	medium	16.7 %	3.7 %
lightning	small	29.2 %	5.0 %
lightning	tiny	23.7 %	8.1 %
lightning	base	52.8 %	8.1 %

Read: WhisperKit medium on auto-detect now matches Lightning — every clip detects its true language (de/es/fr/ja/ko/zh) and transcribes it, where it once mislabelled everything en and translated. zh_001 went 350 % → 0.0 % WER; a 30.8 s Mandarin clip lands at 12.9 % CER (mostly number formatting, 三点 → 3点, not misrecognition). The cause was a config default, not the model: WhisperKit’s DecodingOptions.detectLanguage is false unless you set it, so detection never ran. The other WhisperKit sizes still carry their pre-fix numbers (small 198.6 %, base 284.3 %, tiny 303.3 %, distil-large-v3 231.9 %) — they’ll improve the same way once re-benched; PRs welcome.

Cold start + memory

Engine	Model	Cold start	Peak RSS
lightning	tiny	1.92 s	53 MB
lightning	small	1.10 s	37 MB
lightning	medium	1.17 s	27 MB
lightning	distil-large-v3	1.23 s	23 MB
whisperkit	tiny	6.06 s	103 MB
whisperkit	small	22.41 s	165 MB
whisperkit	medium	27.44 s	197 MB
whisperkit	distil-large-v3	23.12 s	111 MB

Cold-start values include first-run model download. Warm-cache loads are 1–10 s typically.

Recommendation matrix

Memory tier	Default model	Why
8 GB	whisperkit small	4.1 % WER on real speech, 11 % on noisy (acceptable), 165 MB peak — leaves 6+ GB for the agent backend
16 GB (this Mac)	whisperkit medium	2.6 % LibriSpeech WER, 1.3 % multi-voice, 6.3 % noisy. RTF 0.22–0.31× (≥3× realtime), 197 MB. Best balance.
24+ GB	whisperkit medium (same)	No 24 GB host benched yet; revisit when one ships data

Engine default: WhisperKit. Native Swift, zero Python sidecar, faster on Apple Silicon. Lightning stays as the bench baseline.

Multilingual users: auto-detect handles non-English well. Pinning a language in Settings is optional — worth it only to skip the detection pass when you always dictate in one language.

Open questions / next runs

1. large-v3-turbo — neither engine accepts the obvious model name. Discover the right one via WhisperKit.fetchAvailableModels() and rerun. Turbo’s promise is “approaches large-v3 quality at small-ish cost”; if real, it could displace medium as the default.
2. Re-bench the full multilingual matrix post-fix. SPEC-021 re-measured only WhisperKit medium (253 % → 16.7 %). The other sizes (tiny/base/small/distil) still show pre-fix numbers above and should be re-run with detection enabled.
3. Real-world noise types. Synthetic white/pink is a baseline; the M3 corpus should add curated environmental clips (cafe babble, traffic, keyboard) under permissive licensing.
4. Cross-host coverage. 8 GB and 24+ GB tiers are still empty. M1, M2, M3 results all welcome — see bench/CONTRIBUTING.md.
5. Warm-cache cold-start. This run measures cache-miss cold-start. Add a second run that pre-warms each model and reports the warm cold-start (it’s the one users feel after the first launch).

Quality + responsiveness improvements that this data motivates

These feed into M2/M3 specs:

• Pre-warm the chosen model on app launch. 27 s cache-miss cold-start is the worst number; after that it’s ~5–10 s warm. Hide it behind onboarding’s “first dictation test” and the user never sees it.
• VAD-trimmed audio at the recorder boundaries — drop leading/trailing silence before transcribe; smaller window = lower latency.
• User language preference in Settings → General — shipped. Optional, since auto-detect works; pinning a language skips the detection pass for single-language users.
• Custom-words bias. Power users and proper-noun-heavy domains (legal, medical, dev) deserve initial_prompt injection from a user dictionary.
• Streaming. WhisperKit’s transcribeWithResults supports a callback. M3 spec should expose progressive transcripts to the overlay so longer utterances feel snappier.

Reproducing

From the repo root, on the v2 branch:

# Synthetic corpus (multi-voice TTS + multilingual sentences).
bash bench/corpus/fetch.sh

# Real human speech (~337 MB one-time download, openslr.org).
N=20 bash bench/corpus/fetch_librispeech.sh

# Noise-augmented variants (white + pink × 5/10/20 dB SNR).
.venv/bin/python bench/corpus/mix_noise.py --source bench/corpus/voices

# Run the matrix.
swift run openquack-bench \
  --engines whisperkit,lightning \
  --models tiny,base,small,medium,distil-large-v3 \
  --corpus bench/corpus \
  --verbose

Output: bench/out/<host-tag>/{report.md,report.csv,host.json}. Submit yours via PR per bench/CONTRIBUTING.md.

This site is open source. Improve this page.