README.md

metadata

title: GeoTalk
emoji: 🚀
colorFrom: green
colorTo: yellow
sdk: gradio
sdk_version: 6.19.0
python_version: '3.13'
app_file: app.py
pinned: false

Watch AI geoscientists debate your geological data — live, with animated robot avatars!

---

🎬 What is GeoTalk?

GeoTalk is an AI-powered geoscience podcast generator that turns your geological data into an engaging debate between AI scientists — complete with animated 3D robot avatars, text-to-speech voices, and live captions.

Upload a core photo, well log, or outcrop image, and watch two AI geoscientists (plus a moderator) debate the interpretation in real-time.

┌─────────────────────────────────────────────────────────┐
│  🎙️ GeoTalk Live                          ● LIVE       │
├─────────────┬──────────┬────────────────────────────────┤
│             │          │                                │
│   🤖 Robot A │  🎙️ Chair │   🤖 Robot B                   │
│  (Speaker 1) │(Moderator)│  (Speaker 2)                  │
│             │          │                                │
├─────────────┴──────────┴────────────────────────────────┤
│  Dr. Vasquez: "The laminations suggest a tidal flat..." │
├─────────────────────────────────────────────────────────┤
│  ⏸ ════════════════════════ 2:34 / 5:12                │
└─────────────────────────────────────────────────────────┘

✨ Features

Feature	Description
🔬 Core Talk	Two petrographers debate your core/thin section photo
🩺 Log Doctor	Two analysts diagnose your well log data (LAS files) like doctors diagnosing a patient
🏕️ Field Trip FM	A professor and student narrate a virtual field trip to your outcrop
🎙️ Chair/Moderator	A professional host introduces the topic, speakers, and guides the debate
🤖 3D Robot Avatars	Reachy Mini-inspired robots animate per speaker turn (Three.js)
🗣️ Multi-voice TTS	Each character has a distinct voice (Edge-TTS)
📝 Live Captions	Real-time subtitles synced to audio playback
▶️ Auto-play	Episode starts automatically when ready

🏗️ Architecture

┌──────────────┐     ┌──────────────────┐     ┌─────────────┐
│  User Input  │────▶│  Vision Model    │────▶│  LLM Script │
│  (Image/LAS) │     │  (Qwen2.5-VL)   │     │  Generator  │
└──────────────┘     └──────────────────┘     └──────┬──────┘
                                                      │
                     ┌──────────────────┐             │
                     │   Edge-TTS       │◀────────────┘
                     │  (Multi-voice)   │
                     └────────┬─────────┘
                              │
                     ┌────────▼─────────┐
                     │  Robot Viewer    │
                     │  (Three.js +    │
                     │   Audio Sync)   │
                     └─────────────────┘

Pipeline:

1. Vision Analysis — Qwen2.5-VL-72B describes the uploaded geological image
2. Script Generation — Qwen2.5-72B writes a multi-character debate script (with Chair)
3. Text-to-Speech — Edge-TTS generates distinct voices for each character
4. Live Viewer — Three.js renders animated robots synced to audio playback

🚀 Quick Start

Run locally

git clone https://github.com/ShebMichel/GeoTalk.git
cd GeoTalk
pip install -r requirements.txt
export HF_TOKEN="your_huggingface_token"
python app.py

Try the live demo

👉 Launch GeoTalk on Hugging Face Spaces

📁 Project Structure

GeoTalk/
├── app.py              # Gradio UI + tab functions
├── backend.py          # LLM client, TTS generation, script parsing
├── robot_viewer.html   # Three.js robot viewer (self-contained)
├── requirements.txt    # Python dependencies
└── README.md           # This file

🤖 Robot Avatars

The 3D robot avatars are inspired by Reachy Mini from the Small Talk project. They feature:

• Procedural Three.js geometry — no external model files needed
• Per-speaker animations — talking (head bob, antenna wiggle), listening (head tilt), idle (gentle sway)
• Visual feedback — active speaker panel highlighted with colored border
• Chair indicator — center panel with pulsing microphone icon

🛠️ Tech Stack

• Frontend: Three.js, Gradio, HTML5 Audio
• LLM: Qwen2.5-72B-Instruct (script generation)
• Vision: Qwen2.5-VL-72B-Instruct (image analysis)
• TTS: Edge-TTS (Microsoft Neural voices)
• Hosting: Hugging Face Spaces

📄 License

MIT

---

_{Built with ❤️ for the geoscience community}