Spaces:
Runtime error
Runtime error
metadata
title: LFM2-Audio Real-time Speech-to-Speech
emoji: 🎙️
colorFrom: purple
colorTo: pink
sdk: docker
app_port: 7860
pinned: false
license: other
LFM2-Audio Real-time Speech-to-Speech Chat
Real-time WebRTC streaming demo of LFM2-Audio-1.5B, Liquid AI's first end-to-end audio foundation model.
✨ Features
- 🔴 Real-time WebRTC streaming - Instant response with minimal latency
- 🎙️ Continuous listening - Natural conversation flow with automatic pause detection
- 💬 Interleaved output - Simultaneous text and audio generation
- 🔄 Multi-turn memory - Context-aware conversations
- ⚡ Low latency - Optimized for real-time interaction
🚀 How to Use
- Grant microphone access when prompted by your browser
- Start speaking - The model listens continuously
- Pause briefly - The model detects pauses and responds automatically
- Continue conversation - Build multi-turn dialogues naturally
🎛️ Parameters
Temperature
- 0: Greedy decoding (most deterministic)
- 1.0: Default (balanced creativity and coherence)
- 2.0: Maximum creativity (more diverse outputs)
Top-k
- 0: No filtering (full vocabulary)
- 4: Default (conservative, high quality)
- Higher values: More diverse but potentially less coherent
🏗️ Technical Details
- Model: LFM2-Audio-1.5B
- Generation Mode: Interleaved (optimized for real-time)
- Audio Codec: Mimi (24kHz)
- Streaming: WebRTC via fastrtc
- Backend: PyTorch with CUDA acceleration
🔧 Differences from Standard Demo
This demo uses fastrtc for WebRTC streaming, enabling:
- Continuous audio streaming without manual recording
- Automatic voice activity detection (VAD)
- Lower latency through chunked processing
- More natural conversation flow
📚 Resources
📝 License
Licensed under the LFM Open License v1.0
💡 Tips
- Speak clearly and pause briefly between thoughts
- Use a good quality microphone for best results
- Adjust temperature for different creativity levels
- Lower top-k values produce more consistent responses
- GPU acceleration is recommended for real-time performance