🎥 Dolphin: Efficient Audio-Visual Speech Separation with Discrete Lip Semantics and Hierarchical Top-Down Attention

Automatically detect and separate ALL speakers in your video

Simply upload a video and the system will:

  1. 🔍 Automatically detect all speakers in the video
  2. 🎭 Show you each detected speaker's face
  3. 🎬 Generate individual videos for each speaker with their isolated audio

🎬 Try with Example Video
Click to load example video

👇 Click on any face below to view that speaker's video

📖 How it works:

  1. Upload - Select any video file
  2. Process - Click the button to start automatic detection
  3. Review - See all detected speakers and their positions
  4. Select - Click on any face to watch that speaker's separated video

💡 Tips for best results:

  • ✅ Ensure all speakers' faces are visible in the first frame
  • ✅ Use videos with good lighting and clear face views
  • ✅ Works best with frontal or near-frontal face angles
  • ⏱️ Processing time depends on video length and number of speakers

🚀 Powered by:

  • RetinaFace for face detection
  • Dolphin model for audio-visual separation
  • GPU acceleration when available