Cascade - Production-Ready, High-Performance, Asynchronous VAD Library

中文

Cascade is a production-ready, high-performance, and low-latency audio stream processing library designed for Voice Activity Detection (VAD). Built upon the excellent Silero VAD model, Cascade significantly reduces VAD processing latency while maintaining high accuracy through its 1:1:1 binding architecture and asynchronous streaming technology.

📊 Performance Benchmarks

Based on our latest streaming VAD performance tests with different chunk sizes:

Streaming Performance by Chunk Size

Chunk Size (bytes)	Processing Time (ms)	Throughput (chunks/sec)	Total Test Time (s)	Speech Segments
1024	0.66	92.2	3.15	2
4096	1.66	82.4	0.89	2
8192	2.95	72.7	0.51	2

Key Performance Metrics

Metric	Value	Description
Best Processing Speed	0.66ms/chunk	Optimal performance with 1024-byte chunks
Peak Throughput	92.2 chunks/sec	Maximum processing throughput
Success Rate	100%	Processing success rate across all tests
Accuracy	High	Guaranteed by the Silero VAD model
Architecture	1:1:1:1	Independent model per processor instance

Performance Characteristics

• Excellent performance across chunk sizes: High throughput and low latency with various chunk sizes
• Real-time capability: Sub-millisecond processing enables real-time applications
• Scalability: Linear performance scaling with independent processor instances

✨ Core Features

🚀 High-Performance Engineering

• Lock-Free Design: The 1:1:1 binding architecture eliminates lock contention, boosting performance.
• Frame-Aligned Buffer: A highly efficient buffer optimized for 512-sample frames.
• Asynchronous Streaming: Non-blocking audio stream processing based on asyncio.
• Memory Optimization: Zero-copy design, object pooling, and cache alignment.
• Concurrency Optimization: Dedicated threads, asynchronous queues, and batch processing.

🎯 Intelligent Interaction

• Real-time Interruption Detection: VAD-based intelligent interruption detection, allowing users to interrupt system responses at any time
• State Synchronization Guarantee: Two-way guard mechanism ensures strong consistency between physical and logical layers
• Automatic State Management: VAD automatically manages speech collection state, external services control processing state
• Anti-false-trigger Design: Minimum interval checking and state mutex locks effectively prevent false triggers
• Low-latency Response: Interruption detection latency < 50ms for natural conversation experience

🔧 Robust Software Engineering

• Modular Design: A component architecture with high cohesion and low coupling.
• Interface Abstraction: Dependency inversion through interface-based design.
• Type System: Data validation and type checking using Pydantic.
• Comprehensive Testing: Unit, integration, and performance tests.
• Code Standards: Adherence to PEP 8 style guidelines.

🛡️ Production-Ready Reliability

• Error Handling: Robust error handling and recovery mechanisms.
• Resource Management: Automatic cleanup and graceful shutdown.
• Monitoring Metrics: Real-time performance monitoring and statistics.
• Scalability: Horizontal scaling by increasing the number of instances.
• Stability Assurance: Handles boundary conditions and exceptional cases gracefully.

🏗️ Architecture

Cascade employs a 1:1:1:1 independent architecture to ensure optimal performance and thread safety.

graph TD
    Client --> StreamProcessor
    
    subgraph "1:1:1:1 Independent Architecture"
        StreamProcessor --> |per connection| IndependentProcessor[Independent Processor Instance]
        IndependentProcessor --> |independent loading| VADModel[Silero VAD Model]
        IndependentProcessor --> |independent management| VADIterator[VAD Iterator]
        IndependentProcessor --> |independent buffering| FrameBuffer[Frame-Aligned Buffer]
        IndependentProcessor --> |independent state| StateMachine[State Machine]
    end
    
    subgraph "Asynchronous Processing Flow"
        VADModel --> |asyncio.to_thread| VADInference[VAD Inference]
        VADInference --> StateMachine
        StateMachine --> |None| SingleFrame[Single Frame Output]
        StateMachine --> |start| Collecting[Start Collecting]
        StateMachine --> |end| SpeechSegment[Speech Segment Output]
    end

Loading

🚀 Quick Start

Installation

pip install cascade-vad

OR

# Using uv is recommended
uv venv -p 3.12

source .venv/bin/activate

# Install from PyPI (recommended)
pip install cascade-vad

# Or install from source
git clone https://github.com/xucailiang/cascade.git
cd cascade
pip install -e .

Basic Usage

import cascade
import asyncio

async def basic_example():
    """A basic usage example."""
    
    # Method 1: Simple file processing
    async for result in cascade.process_audio_file("audio.wav"):
        if result.result_type == "segment":
            segment = result.segment
            print(f"🎤 Speech Segment: {segment.start_timestamp_ms:.0f}ms - {segment.end_timestamp_ms:.0f}ms")
        else:
            frame = result.frame
            print(f"🔇 Single Frame: {frame.timestamp_ms:.0f}ms")
    
    # Method 2: Stream processing
    async with cascade.StreamProcessor() as processor:
        async for result in processor.process_stream(audio_stream):
            if result.result_type == "segment":
                segment = result.segment
                print(f"🎤 Speech Segment: {segment.start_timestamp_ms:.0f}ms - {segment.end_timestamp_ms:.0f}ms")
            else:
                frame = result.frame
                print(f"🔇 Single Frame: {frame.timestamp_ms:.0f}ms")

asyncio.run(basic_example())

Advanced Configuration

import cascade

async def advanced_example():
    """An advanced configuration example."""
    
    # Custom configuration
    config = cascade.Config(
        vad_threshold=0.7,          # Higher detection threshold
        min_silence_duration_ms=100,
        speech_pad_ms=100
    )
    
    # Use the custom config
    async with cascade.StreamProcessor(config) as processor:
        # Process audio stream
        async for result in processor.process_stream(audio_stream):
            # Process results...
            pass
        
        # Get performance statistics
        stats = processor.get_stats()
        print(f"Throughput: {stats.throughput_chunks_per_second:.1f} chunks/sec")

asyncio.run(advanced_example())

Interruption Detection

import cascade

async def interruption_example():
    """Interruption detection example"""
    
    # Configure interruption detection
    config = cascade.Config(
        vad_threshold=0.5,
        interruption_config=cascade.InterruptionConfig(
            enable_interruption=True,  # Enable interruption detection
            min_interval_ms=500        # Minimum interruption interval 500ms
        )
    )
    
    async with cascade.StreamProcessor(config) as processor:
        async for result in processor.process_stream(audio_stream):
            
            # Detect interruption events
            if result.is_interruption:
                print(f"🛑 Interruption detected! Interrupted state: {result.interruption.system_state.value}")
                # Stop current TTS playback
                await tts_service.stop()
                # Cancel LLM request
                await llm_service.cancel()
            
            # Process speech segments
            elif result.is_speech_segment:
                # ASR recognition
                text = await asr_service.recognize(result.segment.audio_data)
                
                # Set to processing
                processor.set_system_state(cascade.SystemState.PROCESSING)
                
                # LLM generation
                response = await llm_service.generate(text)
                
                # Set to responding
                processor.set_system_state(cascade.SystemState.RESPONDING)
                
                # TTS playback
                await tts_service.play(response)
                
                # Reset to idle after completion
                processor.set_system_state(cascade.SystemState.IDLE)

asyncio.run(interruption_example())

For detailed documentation, see: Interruption Implementation Summary

🧪 Testing

# Run basic integration tests
python tests/test_simple_vad.py -v

# Run simulated audio stream tests
python tests/test_stream_vad.py -v

# Run performance benchmark tests
python tests/benchmark_performance.py

Test Coverage:

• ✅ Basic API Usage
• ✅ Stream Processing
• ✅ File Processing
• ✅ Real Audio VAD
• ✅ Automatic Speech Segment Saving
• ✅ 1:1:1:1 Architecture Validation
• ✅ Performance Benchmarks
• ✅ FrameAlignedBuffer Tests

🌐 Web Demo

We provide a complete WebSocket-based web demonstration that showcases Cascade's real-time VAD capabilities with multiple client support.

Features

• Real-time Audio Processing: Capture audio from browser microphone and process with VAD
• Live VAD Visualization: Real-time display of VAD detection results
• Speech Segment Management: Display detected speech segments with playback support
• Dynamic VAD Configuration: Adjust VAD parameters in real-time
• Multi-client Support: Independent Cascade instances for each WebSocket connection

Quick Start

# Start backend server
cd web_demo
python server.py

# Start frontend (in another terminal)
cd web_demo/frontend
pnpm install && pnpm dev

For detailed setup instructions, see Web Demo Documentation.

🔧 Production Deployment

Best Practices

1. Resource Allocation

   • Each instance uses approximately 50MB of memory.
   • Recommended: 2-3 instances per CPU core.
   • Monitor memory usage to prevent Out-of-Memory (OOM) errors.

2. Performance Tuning

   • Adjust max_instances to match server CPU cores.
   • Increase buffer_size_frames for higher throughput.
   • Tune vad_threshold to balance accuracy and sensitivity.

3. Error Handling

   • Implement retry mechanisms for transient errors.
   • Use health checks to monitor service status.
   • Log detailed information for troubleshooting.

Monitoring Metrics

# Get performance monitoring metrics
stats = processor.get_stats()

# Key monitoring metrics
print(f"Total Chunks Processed: {stats.total_chunks_processed}")
print(f"Average Processing Time: {stats.average_processing_time_ms:.2f}ms")
print(f"Throughput: {stats.throughput_chunks_per_second:.1f} chunks/sec")
print(f"Speech Segments: {stats.speech_segments}")
print(f"Error Rate: {stats.error_rate:.2%}")
print(f"Memory Usage: {stats.memory_usage_mb:.1f}MB")

🔧 Requirements

Core Dependencies

• Python: 3.12 (recommended)
• pydantic: 2.4.0+ (Data validation)
• numpy: 1.24.0+ (Numerical computation)
• scipy: 1.11.0+ (Signal processing)
• silero-vad: 5.1.2+ (VAD model)
• onnxruntime: 1.22.1+ (ONNX inference)
• torchaudio: 2.7.1+ (Audio processing)

Development Dependencies

• pytest: Testing framework
• black: Code formatter
• ruff: Linter
• mypy: Type checker
• pre-commit: Git hooks

🤝 Contribution Guide

We welcome community contributions! Please follow these steps:

1. Fork the project and create a feature branch.
2. Install development dependencies: pip install -e .[dev]
3. Run tests: pytest
4. Lint your code: ruff check . && black --check .
5. Type check: mypy cascade
6. Submit a Pull Request with a clear description of your changes.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• Silero Team: For their excellent VAD model.
• PyTorch Team: For the deep learning framework.
• Pydantic Team: For the type validation system.
• Python Community: For the rich ecosystem.

📞 Contact

• Author: Xucailiang
• Email: xucailiang.ai@gmail.com
• Project Homepage: https://github.com/xucailiang/cascade
• Issue Tracker: https://github.com/xucailiang/cascade/issues
• Documentation: https://cascade-vad.readthedocs.io/

---

⭐ If you find this project helpful, please give it a star!