Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →OpenAI Whisper — general-purpose speech recognition. Multilingual transcription, translation to English, and speaker-agnostic ASR. Models from tiny to large. Robust to noise, accents, and technical vocabulary.
.claude/skills/mkurman-whisper/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-07 | ✗→✓ | ▲ Improved | -39% | 0% |
| case-18 | ✗→✓ | ▲ Improved | 30% | 0% |
| case-15 | ✓→✓ | ＝ Same ✓ | -38% | 0% |
| case-01 | ✓→✓ | ＝ Same ✓ | 7% | 0% |
| case-02 | ✓→✓ | ＝ Same ✓ | -1% | 0% |
OpenAI Whisper is a general-purpose speech recognition model supporting multilingual transcription, translation to English, and speaker-agnostic ASR. Models range from tiny (39M params) to large (1.55B params). Robust to noise, accents, and technical vocabulary.
bashuv pip install openai-whisper ffmpeg # required for audio loading
pythonimport whisper model = whisper.load_model("base") result = model.transcribe("audio.mp3") print(result["text"])
python# Transcribe in original language result = model.transcribe("french_audio.mp3", language="fr") # Translate to English result = model.transcribe("german_audio.mp3", task="translate") print(result["text"]) # English output
pythonresult = model.transcribe("lecture.mp3", word_timestamps=True) for segment in result["segments"]: for word in segment["words"]: print(f"{word['word']}: {word['start']:.2f}-{word['end']:.2f}")
python# tiny (fast, less accurate) → base → small → medium → large (slow, most accurate) sizes = ["tiny", "base", "small", "medium", "large"] for s in sizes: m = whisper.load_model(s) # ~1GB VRAM for base, ~10GB for large result = m.transcribe("podcast.mp3")
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-15 | pass→pass | 7,978 | 2,525 | -68% | 1 | 1 | 0% | 1,277 | 790 | -38% | 0 | 0 | — |
case-01 | pass→pass | 5,149 | 3,704 | -28% | 1 | 1 | 0% | 974 | 1,038 | +7% | 0 | 0 | — |
case-02 | pass→pass | 5,476 | 3,155 | -42% | 1 | 1 | 0% | 945 | 936 | -1% | 0 | 0 | — |
case-03 | pass→pass | 6,654 | 1,495 | -78% | 1 | 1 | 0% | 1,111 | 654 | -41% | 0 | 0 | — |
case-04 | pass→pass | 4,842 | 1,449 | -70% | 1 | 1 | 0% | 774 | 666 | -14% | 0 | 0 | — |
case-05 | pass→pass | 2,533 | 2,248 | -11% | 1 | 1 | 0% | 364 | 780 | +114% | 0 | 0 | — |
case-06 | pass→pass | 6,631 | 2,457 | -63% | 1 | 1 | 0% | 1,268 | 789 | -38% | 0 | 0 | — |
case-07 | fail→pass | 9,121 | 3,709 | -59% | 1 | 1 | 0% | 1,706 | 1,034 | -39% | 0 | 0 | — |
case-08 | pass→pass | 9,445 | 6,224 | -34% | 1 | 1 | 0% | 1,507 | 1,466 | -3% | 0 | 0 | — |
case-09 | pass→pass | 11,554 | 5,646 | -51% | 1 | 1 | 0% | 2,149 | 1,413 | -34% | 0 | 0 | — |
case-10 | pass→pass | 7,061 | 3,351 | -53% | 1 | 1 | 0% | 1,261 | 962 | -24% | 0 | 0 | — |
case-11 | pass→pass | 6,428 | 3,753 | -42% | 1 | 1 | 0% | 1,180 | 1,129 | -4% | 0 | 0 | — |
case-12 | pass→pass | 9,377 | 7,571 | -19% | 1 | 1 | 0% | 1,534 | 1,801 | +17% | 0 | 0 | — |
case-13 | pass→pass | 9,805 | 1,558 | -84% | 1 | 1 | 0% | 1,872 | 619 | -67% | 0 | 0 | — |
case-14 | pass→pass | 8,680 | 1,515 | -83% | 1 | 1 | 0% | 1,455 | 617 | -58% | 0 | 0 | — |
case-16 | pass→pass | 9,734 | 6,477 | -33% | 1 | 1 | 0% | 1,610 | 1,541 | -4% | 0 | 0 | — |
case-17 | pass→pass | 11,981 | 7,997 | -33% | 1 | 1 | 0% | 1,979 | 1,734 | -12% | 0 | 0 | — |
case-18 | fail→pass | 3,129 | 1,661 | -47% | 1 | 1 | 0% | 482 | 625 | +30% | 0 | 0 | — |
case-19 | pass→pass | 2,647 | 2,230 | -16% | 1 | 1 | 0% | 455 | 768 | +69% | 0 | 0 | — |
case-20 | pass→pass | 11,864 | 10,010 | -16% | 1 | 1 | 0% | 2,099 | 2,319 | +10% | 0 | 0 | — |
case-21 | pass→pass | 7,503 | 7,911 | +5% | 1 | 1 | 0% | 1,229 | 1,944 | +58% | 0 | 0 | — |
case-22 | pass→pass | 6,439 | 5,724 | -11% | 1 | 1 | 0% | 1,233 | 1,445 | +17% | 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 +9 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.