Features are continuously being updated. Stay tuned for our latest developments... ✨

Fast-VC-Service aims to build a high-performance real-time streaming voice conversion cloud service designed for production environments. Based on the Seed-VC model, it supports WebSocket protocol and PCM/OPUS audio encoding formats.

✨ Core Features

• Real-time Conversion: Low-latency streaming voice conversion based on Seed-VC
• WebSocket API: Support for PCM and OPUS audio formats
• Performance Monitoring: Complete real-time performance metrics statistics
• High Concurrency: Multi-Worker concurrent processing, supporting production environments
• Easy Deployment: Simple configuration, one-click startup

🚀 Quick Start

📦 One-click Installation

# Install system dependencies (Ubuntu/Debian)
sudo apt-get update
sudo apt-get install -y libopus-dev libopus0 opus-tools

# Clone project
git clone --recursive https://github.com/Leroll/fast-vc-service.git
cd fast-vc-service

# Configure environment
cp .env.example .env

# Install dependencies (using uv)
uv sync

# Start service
uv run fast-vc serve

🧪 Quick Testing

# WebSocket real-time voice conversion
uv run examples/websocket/ws_client.py

For detailed installation and usage guide, please refer to Quick Start documentation.

📈 Performance

• Time unit: milliseconds (ms)
• View detailed test report:
  • Performance-Report_4090D
  • Performance-Report_1080ti

📝 Version Updates

2025-11-04 - v0.1.7: Audio quality improvements and output sample-rate control

• Quality
  • Added t_span_schedule (supports cosine reordering) to improve quality at the same step count
  • Unified realtime_vc output to 22 kHz for better fidelity
  • Preserve sufficient voiced context across long silences to improve coherence
• Output sample rate
  • WebSocket adds sample_rate_out to return audio at the requested sample rate
• Other
  • VAD settings moved into config for easier tuning and portability
  • time_records centralized under session management
  • Removed realtimevc perf tracing, file-vc scripts, and noise-gate module to reduce redundancy
  • Fixed audio not being persisted when upstream disconnects unexpectedly

2025-10-16 - v0.1.6: Multi-GPU/Multi-instance deployment, semantic feature retrieval, uv package management

• Deployment & Concurrency
  • Support custom model and config file paths, enabling multiple instances with isolated configs and ports
  • Support single instance using multiple GPUs to improve concurrency
• Quality & Effect
  • Added semantic feature retrieval module to improve voice similarity and robustness
  • Optimized VAD parameters to reduce noise and silence false triggers
• Engineering
  • Migrated package management to uv for faster install and startup
  • Fixed send_slow false delay warning
  • Added VC evaluation tool tools/eval.py

2025-07-24 - v0.1.5: Pitch Adaptive Matching Support and Real-time Monitoring Optimization

• Real-time monitoring optimization:
  • Optimized timeline_lognize, added delay items statistics for same event types
  • Added SLOW tags to logs for monitoring receive intervals, send intervals, and VC-E2E latency
• Support for pitch adaptive matching with reference audio to improve conversion quality
  • Added pitch analysis script providing audio analysis tools
  • Added pitch adaptive matching functionality with corresponding toggle configuration
• Other optimizations
  • Changed UID generation method to time-based generation for easier experimentation and testing
  • Optimized session tool's file naming mechanism
  • Added config and model path options, support NAS configuration files, enable simpler cloud host deployment

🚧 TODO

• tag - v0.2 - Improve inference efficiency, reduce RTF
  • Add a monitoring frontend to display performance metrics and system status in real time
  • Add scheduled tasks to periodically clean up logs and persisted data
  • Optimize timeline analysis, add batch analysis to the comprehensive report, and add a streaming smoothness metric to determine whether returned audio is sufficient for downstream playback
  • Support fp16 inference mode
  • Update documentation to reflect the latest code
  • Train models to optimize voice conversion quality
  • Improve model effectiveness for noisy data
    • Distinguish different noise types
  • Server send/recv event definitions should match roles
  • Model acceleration optimization:
    • Change VAD to use ONNX-GPU to improve inference speed
    • Explore solutions to reduce model inference latency (e.g., new model architectures, quantization, etc.)
  • Create Docker image and AutoDL image for one-click deployment

🙏 Acknowledgements

• Seed-VC - Provides powerful underlying voice conversion model
• RVC - Provides basic streaming voice conversion pipeline