Whisper-Transcription

npx skills add majiayu000/claude-skill-registry --skill whisper-transcription --agent claude-code

Same command for any agent — swap --agent for codex, cursor, copilot.

# Whisper Transcription Server Local speech-to-text transcription using whisper.cpp with GPU acceleration. The server runs on port 5555 and accepts audio files via HTTP POST. ## Instructions When helping users with audio transcription, follow these guidelines: 1. **Server Location**: The whisper-server runs at `http://localhost:5555` with the large-v3 model 2. **Audio Format**: Server accepts WAV, MP3, and other common formats. 16kHz mono WAV is optimal 3. **API Endpoint**: Use POST to `/inference` with multipart form data 4. **GPU Memory**: large-v3 uses ~6GB VRAM on the RTX 3090 (24GB total) 5. **VRAM Sharing**: If OOM errors occur, wait 30-60 seconds for other GPU services to unload (see Vram-GPU-OOM skill) ## Examples ### Example 1: Basic Transcription ``` User: Transcribe this audio file Claude: I'll send the audio to the local Whisper server: curl -X POST http://localhost:5555/inference \ -F "file=@audio.wav" The server returns JSON with the transcribed text. ``` ### Example 2: Recording and Transcribing ``` User: Record my voice and transcribe it Claude: I'll record audio then send it to Whisper: 1. Record 5 seconds of audio: arecord -d 5 -f S16_LE -r 16000 -c 1 recording.wa

curl -X POST http://localhost:5555/inference \
transcribe.sh - Quick transcription helper
if [ -z "$1" ]; then
echo "Usage: transcribe.sh <audio_file>"
exit 1
fi
curl -s -X POST "$WHISPER_URL/inference" \
Check model exists
ls -lh ~/whisper.cpp/models/ggml-large-v3.bin
Check CUDA