metadata-generation

name	metadata-generation
description	Generate show metadata (JSON) containing timecoded transcript, title, duration, summary, and date by sending audio and transcript to Gemini.

Metadata Generation

Generate a structured JSON file containing metadata for the radio show. This is done by sending the final mixed audio file and the original transcript back to Gemini.

Prerequisites

pip install google-genai

Workflow

1. Read Inputs: Load the transcript from ./workspace/data/script.md and the final audio from ./workspace/audio/final/ai_radio.mp3.
2. Upload Audio: Upload the audio file to the Gemini Files API.
3. Generate Metadata: Call Gemini (e.g., gemini-3-flash-preview) using the Interactions API, passing the uploaded audio file and the transcript.
4. Save Output: Save the resulting JSON to ./workspace/data/show_notes.json.

Python Code

The skill is implemented in scripts/generate_metadata.py.

# Example usage
python3 skills/metadata-generation/scripts/generate_metadata.py --workspace ./workspace

Output

The skill produces a JSON file at ./workspace/data/show_notes.json with the following structure:

{
  "show_title": "...",
  "show_duration": "...",
  "two_sentence_summary": "...",
  "date_of_generation": "YYYY-MM-DD",
  "timecoded_transcript": [
    {
      "timecode": "MM:SS",
      "speaker": "...",
      "text": "..."
    },
    ...
  ]
}

Error Handling

• If the audio file is missing, the script will fail.
• If the transcript is missing, it will fail.
• If the Gemini API call fails, it will report the error and exit.