Whisper Setup

Whisper Setup Skill

This skill guides users through setting up Whisper for the PTT plugin.

Overview

The PTT plugin supports two transcription backends:

1. OpenAI Whisper API - Cloud-based, easy setup, costs ~$0.006/min
2. Local whisper.cpp - Free, private, works offline, requires setup

Setup Flow

Step 1: Ask User Preference

Ask the user which mode they prefer:

Which Whisper mode would you like to set up?

1. **OpenAI API** (Recommended for ease of use)
   - Requires OpenAI API key
   - Costs ~$0.006 per minute of audio
   - Best transcription quality
   - Requires internet connection

2. **Local whisper.cpp** (Recommended for privacy)
   - Free, no API costs
   - Works offline
   - Requires ~150MB-3GB disk space (depending on model)
   - Transcription speed depends on your hardware

Step 2A: OpenAI API Setup

If user chooses API:

1. Check if OPENAI_API_KEY environment variable is set:

   echo $OPENAI_API_KEY | head -c 10
   

2. If not set, ask user to provide their API key

3. Update config:

   # Read current config and update
   cat ~/.claude/ptt-config.json
   

   Set whisper.openaiApiKey to the user's key or instruct them to set OPENAI_API_KEY env var.

4. Set whisper.preferredMode to "api"

Step 2B: Local whisper.cpp Setup

If user chooses local:

Check System Resources

1. Check available RAM:

   free -h
   

2. Check available disk space:

   df -h ~
   

3. Check CPU info:

   lscpu | grep -E "(Model name|CPU\(s\)|Thread)"
   

4. Check for NVIDIA GPU (for CUDA acceleration):

   nvidia-smi 2>/dev/null || echo "No NVIDIA GPU detected"
   

Recommend Model Based on Resources

Present model options with recommendations based on system:

Recommendations:

• RAM < 4GB: Use tiny.en
• RAM 4-8GB: Use base.en (default recommendation)
• RAM 8-16GB: Use small.en for better quality
• RAM > 16GB: Use medium.en or large-v3 if accuracy is critical

Install whisper.cpp

1. Install dependencies:

   sudo apt-get update && sudo apt-get install -y build-essential cmake
   

2. Clone and build:

   cd ~ && git clone https://github.com/ggerganov/whisper.cpp.git
   cd whisper.cpp && make -j$(nproc)
   

3. Download chosen model:

   ./models/download-ggml-model.sh <model_name>
   

   Replace <model_name> with: tiny.en, base.en, small.en, medium.en, or large-v3

4. Test the installation:

   ./build/bin/whisper-cli -m models/ggml-<model>.bin -f samples/jfk.wav
   

Update Config

Update ~/.claude/ptt-config.json:

{
  "whisper": {
    "localModelPath": "/home/<user>/whisper.cpp/models/ggml-<model>.bin",
    "whisperExecutable": "/home/<user>/whisper.cpp/build/bin/whisper-cli",
    "preferredMode": "local"
  }
}

Step 3: Test Configuration

Verify the setup works:

1. For API mode, the MCP server should be able to make API calls
2. For local mode, test whisper-cli directly:

   # Record a short test
   arecord -f S16_LE -r 16000 -c 1 -d 3 /tmp/test.wav
   
   # Transcribe
   ~/whisper.cpp/build/bin/whisper-cli -m ~/whisper.cpp/models/ggml-base.en.bin -f /tmp/test.wav
   

Step 4: Enable Fallback (Optional)

Ask if user wants fallback enabled:

• If both API key and local model are configured, enable enableFallback: true
• This provides resilience - if one method fails, the other is tried

Troubleshooting

Common Issues

1. "whisper-cli not found"

   • Ensure whisper.cpp was built successfully
   • Check the executable path in config

2. "Model file not found"

   • Verify the model was downloaded
   • Check the model path in config

3. "API key invalid"

   • Verify the API key is correct
   • Check for extra whitespace

4. "Out of memory" during local transcription

   • Use a smaller model
   • Close other applications

Platform-Specific Notes

macOS

• whisper.cpp builds with CoreML acceleration on Apple Silicon
• Use make without additional flags

Linux with NVIDIA GPU

• Build with CUDA support:

  make GGML_CUDA=1
  

Windows

• Use WSL2 for best compatibility
• Or download pre-built binaries from whisper.cpp releases