A "fan-made" AI-Powered LLM-based editorial tool to process and review radio show archive. It uses Google's Gemini models (multimodal) to transcribe audio, detect music segments, and generate structured metadata (summaries, hashtags, categories).
- Install dependencies:
pip install -r requirements.txt
- Set API Key:
export GEMINI_API_KEY="your_gemini_api_key"
- Run the app:
streamlit run app.py
- Direct Upload: Audio files (m4a) are uploaded directly to Gemini's multimodal context window.
- Smart Transcription:
- Distinguishes spoken content from music.
- Extracts music segments with timestamps (
[MUSIC PLAYING: start - end]). - Generate AI summary and identifies tone and vibe tags.
- Configurable Prompts: Driven by
prompts.tomlto iterate without code changes.
- Status Workflow: Track episodes through
Pending→Approved/Rejected. - Toast Notifications: Custom persistent notification queue for user feedback.
- Statistics: Real-time metrics on repository status.
- Audio Player: with a visualized music timeline and a “Jump to next song” button.
- Dual-Column View: Read-only transcript (left) vs. Editable summary (right).
- Tagging System: Multi-select UI for Categories, Tags, Tone, and Vibe.
- Diff Viewer: Visual highlights of changes between original AI generation and editorial edits.
- Baseline Comparison: Compare current summaries against historical baselines (via CLI flag
--base).
- Frontend: Streamlit (Python)
- AI Engine: Google Gemini (Multimodal) GenAI SDK (
gemini-pro-latest) - Datastore: JSON file (
data/episodes.json) - Audio Handling:
mutagen(metadata), Gemini File API - Prompt Engineering: Externalized
prompts.toml
1. Missing Dependencies
- Ensure you are running Python 3.11+ and have installed all packages:
pip install -r requirements.txt.
2. Audio Upload Failures
- Ensure
GEMINI_API_KEYis set correctly and has permissions for the Gemini File API.
3. API Rate Limits (429 Errors)
- Check your Google Cloud project quotas for the model you’re using (e.g.,
gemini-pro-latest).
