Gemini Speech-to-Text Skill

Gemini Speech-to-Text Skill

Transcribe audio files using Google's Gemini API or Vertex AI. Default model is gemini-2.0-flash-lite for fastest transcription.

Authentication (choose one)

Option 1: Vertex AI with Application Default Credentials (Recommended)

gcloud auth application-default login
gcloud config set project YOUR_PROJECT_ID

The script will automatically detect and use ADC when available.

Option 2: Direct Gemini API Key

Set GEMINI_API_KEY in environment (e.g., ~/​.env or ~/​.clawdbot/​.env)

Requirements

• Python 3.10+ (no external dependencies)
• Either GEMINI_API_KEY or gcloud CLI with ADC configured

Supported Formats

• .ogg / .opus (Telegram voice messages)
• .mp3
• .wav
• .m4a

Usage

# Auto-detect auth (tries ADC first, then GEMINI_API_KEY)
python ~/​.claude/​skills/​gemini-stt/​transcribe.py /​path/​to/​audio.ogg

# Force Vertex AI
python ~/​.claude/​skills/​gemini-stt/​transcribe.py /​path/​to/​audio.ogg --vertex

# With a specific model
python ~/​.claude/​skills/​gemini-stt/​transcribe.py /​path/​to/​audio.ogg --model gemini-2.5-pro

# Vertex AI with specific project and region
python ~/​.claude/​skills/​gemini-stt/​transcribe.py /​path/​to/​audio.ogg --vertex --project my-project --region us-central1

# With Clawdbot media
python ~/​.claude/​skills/​gemini-stt/​transcribe.py ~/​.clawdbot/​media/​inbound/​voice-message.ogg

Options

Supported Models

Any Gemini model that supports audio input can be used. Recommended models:

See Gemini API Models for the latest list.

How It Works

1. Reads the audio file and base64 encodes it
2. Auto-detects authentication:
   • If ADC is available (gcloud), uses Vertex AI endpoint
   • Otherwise, uses GEMINI_API_KEY with direct Gemini API
3. Sends to the selected Gemini model with transcription prompt
4. Returns the transcribed text

Example Integration

For Clawdbot voice message handling:

# Transcribe incoming voice message
TRANSCRIPT=$(python ~/​.claude/​skills/​gemini-stt/​transcribe.py "$AUDIO_PATH")
echo "User said: $TRANSCRIPT"

Error Handling

The script exits with code 1 and prints to stderr on:

• No authentication available (neither ADC nor GEMINI_API_KEY)
• File not found
• API errors
• Missing GCP project (when using Vertex)

Notes

• Uses Gemini 2.0 Flash Lite by default for fastest transcription
• No external Python dependencies (uses stdlib only)
• Automatically detects MIME type from file extension
• Prefers Vertex AI with ADC when available (no API key management needed)