Transcribe audio files via OpenRouter using audio-capable models

Transcribe audio files using OpenRouter's chat completions API with input_audio content type. Works with any audio-capable model.

Quick start

{baseDir}/scripts/transcribe.sh /path/to/audio.m4a

Output goes to stdout.

Useful flags

# Custom model (default: google/gemini-2.5-flash)
{baseDir}/scripts/transcribe.sh audio.ogg --model openai/gpt-4o-audio-preview

# Custom instructions
{baseDir}/scripts/transcribe.sh audio.m4a --prompt "Transcribe with speaker labels"

# Save to file
{baseDir}/scripts/transcribe.sh audio.m4a --out /tmp/transcript.txt

# Custom caller identifier (for OpenRouter dashboard)
{baseDir}/scripts/transcribe.sh audio.m4a --title "MyApp"

How it works

1. Converts audio to WAV (mono, 16kHz) using ffmpeg
2. Base64 encodes the audio
3. Sends to OpenRouter chat completions with input_audio content
4. Extracts transcript from response

API key

Set OPENROUTER_API_KEY env var, or configure in ~/.clawdbot/clawdbot.json:

{
  skills: {
    "openrouter-transcribe": {
      apiKey: "YOUR_OPENROUTER_KEY"
    }
  }
}

Headers

The script sends identification headers to OpenRouter:

1. X-Title: Caller name (default: "Peanut/Clawdbot")
2. HTTP-Referer: Reference URL (default: "https://clawdbot.com")

These show up in your OpenRouter dashboard for tracking.

Troubleshooting

ffmpeg format errors: The script uses a temp directory (not mktemp -t file.wav) because macOS's mktemp adds random suffixes after the extension, breaking format detection.

Argument list too long: Large audio files produce huge base64 strings that exceed shell argument limits. The script writes to temp files (--rawfile for jq, @file for curl) instead of passing data as arguments.

Empty response: If you get "Empty response from API", the script will dump the raw response for debugging. Common causes:

1. Invalid API key
2. Model doesn't support audio input
3. Audio file too large or corrupted