Loading skill
Install any skill in seconds. Free to start, no credit card required.
Get Started Free →SpeechBrain — PyTorch speech toolkit. ASR, speaker recognition, speech separation, diarization, enhancement, language identification, and TTS. Recipe-based training with pre-trained model zoo.
.claude/skills/mkurman-speechbrain/SKILL.md| Test case | Without → With | Effect | Δ tokens | Δ turns |
|---|---|---|---|---|
| case-01 | ✗→✓ | ▲ Improved | -46% | 0% |
| case-02 | ✗→✓ | ▲ Improved | -32% | 0% |
| case-03 | ✗→✓ | ▲ Improved | -11% | 0% |
| case-10 | ✗→✓ | ▲ Improved | -43% | 0% |
| case-17 | ✗→✓ | ▲ Improved | -52% | 0% |
SpeechBrain is an open-source PyTorch speech processing toolkit covering ASR (speech-to-text), speaker recognition, speech separation, diarization, enhancement, language identification, emotion recognition, and text-to-speech. Provides pretrained models and recipe-based training.
bashuv pip install speechbrain
pythonfrom speechbrain.inference.ASR import EncoderDecoderASR asr_model = EncoderDecoderASR.from_hparams( source="speechbrain/asr-crdnn-rnnlm-librispeech", savedir="pretrained_models/asr") transcript = asr_model.transcribe_file("audio.wav") print(f"Transcript: {transcript}")
pythonfrom speechbrain.inference.speaker import SpeakerRecognition verification = SpeakerRecognition.from_hparams( source="speechbrain/spkrec-ecapa-voxceleb", savedir="pretrained_models/spkrec") score, prediction = verification.verify_files("speaker1.wav", "speaker2.wav") print(f"Same speaker: {prediction} (score: {score:.3f})")
| Case | Status | Duration (ms) | Turns | Tokens | Tool calls | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Without | With | Δ | Without | With | Δ | Without | With | Δ | Without | With | Δ | ||
case-21 | pass→pass | 16,122 | 10,328 | -36% | 1 | 1 | 0% | 3,121 | 2,285 | -27% | 0 | 0 | — |
case-22 | pass→pass | 11,032 | 7,523 | -32% | 1 | 1 | 0% | 2,010 | 1,700 | -15% | 0 | 0 | — |
case-01 | fail→pass | 7,390 | 2,486 | -66% | 1 | 1 | 0% | 1,387 | 753 | -46% | 0 | 0 | — |
case-02 | fail→pass | 7,685 | 3,740 | -51% | 1 | 1 | 0% | 1,273 | 864 | -32% | 0 | 0 | — |
case-03 | fail→pass | 7,616 | 3,666 | -52% | 1 | 1 | 0% | 951 | 843 | -11% | 0 | 0 | — |
case-04 | pass→pass | 4,330 | 1,767 | -59% | 1 | 1 | 0% | 722 | 623 | -14% | 0 | 0 | — |
case-10 | fail→pass | 5,957 | 2,240 | -62% | 1 | 1 | 0% | 1,067 | 612 | -43% | 0 | 0 | — |
case-05 | fail→fail | 6,544 | 7,863 | +20% | 1 | 1 | 0% | 1,075 | 1,112 | +3% | 0 | 0 | — |
case-06 | pass→pass | 6,656 | 2,875 | -57% | 1 | 1 | 0% | 1,088 | 789 | -27% | 0 | 0 | — |
case-07 | fail→fail | 8,552 | 4,826 | -44% | 1 | 1 | 0% | 1,610 | 1,269 | -21% | 0 | 0 | — |
case-08 | pass→pass | 5,256 | 1,671 | -68% | 1 | 1 | 0% | 936 | 530 | -43% | 0 | 0 | — |
case-09 | pass→pass | 2,911 | 2,473 | -15% | 1 | 1 | 0% | 459 | 737 | +61% | 0 | 0 | — |
case-11 | pass→pass | 4,984 | 2,312 | -54% | 1 | 1 | 0% | 738 | 685 | -7% | 0 | 0 | — |
case-12 | pass→pass | 4,406 | 2,612 | -41% | 1 | 1 | 0% | 713 | 759 | +6% | 0 | 0 | — |
case-13 | fail→fail | 9,007 | 4,747 | -47% | 1 | 1 | 0% | 1,796 | 1,216 | -32% | 0 | 0 | — |
case-14 | pass→pass | 5,690 | 3,572 | -37% | 1 | 1 | 0% | 1,000 | 909 | -9% | 0 | 0 | — |
case-15 | pass→pass | 4,537 | 1,890 | -58% | 1 | 1 | 0% | 788 | 608 | -23% | 0 | 0 | — |
case-16 | fail→fail | 7,571 | 4,378 | -42% | 1 | 1 | 0% | 1,370 | 1,151 | -16% | 0 | 0 | — |
case-17 | fail→pass | 9,846 | 3,362 | -66% | 1 | 1 | 0% | 1,752 | 846 | -52% | 0 | 0 | — |
case-18 | pass→pass | 10,361 | 2,954 | -71% | 1 | 1 | 0% | 1,668 | 864 | -48% | 0 | 0 | — |
case-19 | pass→pass | 8,085 | 4,548 | -44% | 1 | 1 | 0% | 1,403 | 1,002 | -29% | 0 | 0 | — |
case-20 | pass→pass | 18,394 | 16,322 | -11% | 1 | 1 | 0% | 3,608 | 3,859 | +7% | 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 +23 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.