Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →Local speech-to-text with Parakeet MLX (ASR) for Apple Silicon (no API key).
.claude/skills/sundial-org-parakeet-mlx/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-20 | ✗→✓ | ▲ Improved | -30% | 0% |
| case-01 | ✗→✓ | ▲ Improved | -60% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -55% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -42% | 0% |
| case-04 | ✗→✓ | ▲ Improved | -48% | 0% |
Use parakeet-mlx to transcribe audio locally on Apple Silicon.
Quick start
parakeet-mlx /path/audio.mp3 --output-format txtparakeet-mlx /path/audio.m4a --output-format vtt --highlight-wordsparakeet-mlx *.mp3 --output-format allNotes
uv tool install parakeet-mlx -U (not uv add or pip install)parakeet-mlx --help to see all options (--help, not -h).~/.cache/huggingface on first run.mlx-community/parakeet-tdt-0.6b-v3 (optimized for Apple Silicon).ffmpeg installed for audio processing.--verbose for detailed progress and confidence scores.*.mp3 work).| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-20 | fail→pass | 17,380 | 11,378 | -35% | 1 | 1 | 0% | 3,206 | 2,260 | -30% | 0 | 0 | — |
case-01 | fail→pass | 10,916 | 3,014 | -72% | 1 | 1 | 0% | 2,069 | 821 | -60% | 0 | 0 | — |
case-02 | fail→pass | 7,905 | 2,449 | -69% | 1 | 1 | 0% | 1,471 | 667 | -55% | 0 | 0 | — |
case-03 | fail→pass | 6,381 | 2,166 | -66% | 1 | 1 | 0% | 1,150 | 671 | -42% | 0 | 0 | — |
case-04 | fail→pass | 6,524 | 1,894 | -71% | 1 | 1 | 0% | 1,165 | 609 | -48% | 0 | 0 | — |
case-05 | pass→pass | 5,993 | 2,004 | -67% | 1 | 1 | 0% | 1,064 | 626 | -41% | 0 | 0 | — |
case-06 | fail→pass | 8,488 | 1,982 | -77% | 1 | 1 | 0% | 1,620 | 637 | -61% | 0 | 0 | — |
case-07 | pass→pass | 2,809 | 1,675 | -40% | 1 | 1 | 0% | 422 | 502 | +19% | 0 | 0 | — |
case-08 | fail→pass | 8,391 | 1,707 | -80% | 1 | 1 | 0% | 1,391 | 522 | -62% | 0 | 0 | — |
case-09 | pass→pass | 14,744 | 2,582 | -82% | 1 | 1 | 0% | 2,491 | 638 | -74% | 0 | 0 | — |
case-10 | pass→pass | 9,572 | 3,209 | -66% | 1 | 1 | 0% | 1,460 | 813 | -44% | 0 | 0 | — |
case-11 | pass→pass | 9,213 | 2,548 | -72% | 1 | 1 | 0% | 1,521 | 643 | -58% | 0 | 0 | — |
case-12 | pass→pass | 6,261 | 1,722 | -72% | 1 | 1 | 0% | 1,079 | 487 | -55% | 0 | 0 | — |
case-13 | pass→pass | 6,192 | 2,304 | -63% | 1 | 1 | 0% | 1,078 | 597 | -45% | 0 | 0 | — |
case-14 | pass→pass | 8,191 | 2,276 | -72% | 1 | 1 | 0% | 1,401 | 628 | -55% | 0 | 0 | — |
case-15 | pass→pass | 9,576 | 2,467 | -74% | 1 | 1 | 0% | 1,569 | 619 | -61% | 0 | 0 | — |
case-16 | pass→pass | 8,071 | 2,154 | -73% | 1 | 1 | 0% | 1,276 | 601 | -53% | 0 | 0 | — |
case-17 | fail→pass | 4,288 | 1,439 | -66% | 1 | 1 | 0% | 663 | 432 | -35% | 0 | 0 | — |
case-18 | fail→pass | 7,477 | 1,499 | -80% | 1 | 1 | 0% | 1,236 | 475 | -62% | 0 | 0 | — |
case-19 | fail→pass | 17,496 | 8,389 | -52% | 1 | 1 | 0% | 2,832 | 1,615 | -43% | 0 | 0 | — |
case-21 | pass→pass | 8,437 | 4,788 | -43% | 1 | 1 | 0% | 1,591 | 1,068 | -33% | 0 | 0 | — |
case-22 | fail→pass | 4,770 | 1,816 | -62% | 1 | 1 | 0% | 744 | 542 | -27% | 0 | 0 | — |
DecimalAI ran this skill against gemini-3.6-flash twice over the same eval suite — once with the skill loaded and once without — and compared the two runs case by case. 22 cases were attempted. The headline lift of +50 percentage points is the difference between those two pass rates over the 22 comparable cases.
Without the skill loaded, the model failed this case. With it loaded, the same prompt on the same model passed. This is one improved case from the latest verified run; every case, including any that regressed, is in the table above.
Other measured skills in the registry, with their headline benchmark lift.