Wakeword Training Platform

Production-ready platform for training custom wakeword detection models with GPU acceleration, advanced optimizations, and a modern web interface. Features enterprise-grade Distributed Cascade Architecture for real-time deployment.

🚀 Current Version: v4.0 - Production Release
🔧 GPU Support: CUDA 11.8+ with Mixed Precision
🌐 Deployment: ONNX, TensorFlow Lite, Raspberry Pi

---

📚 Quick Navigation

📖 Documentation	🔧 Configuration	🎯 Usage
📘 Complete Guide	⚙️ Presets	🚀 Quick Start
User Guide & Reference	GPU/RPi Optimization	Training & Deployment

🔍 Need help? Check our Technical Features Guide for CMVN, EMA, and FAH metrics.

---

🚀 Quick Start

Prerequisites

• Python: 3.10+
• CUDA: 11.8+ (for GPU acceleration)
• GPU: NVIDIA GPU with 6GB+ VRAM recommended

Installation

1. Clone the Repository

   git clone https://github.com/sarpel/wakeword-training-platform.git
   cd wakeword-training-platform

2. Install Dependencies

   pip install -r requirements.txt

   Note: For PyTorch with CUDA 11.8, see DOCUMENTATION.md.

3. Launch the Application

   python run.py

   The application will open at http://localhost:7860

🚀 Quick Start (Docker - Recommended)

For a consistent environment across Windows and Linux:

1. Configure Environment

   cp .env.example .env
   # Edit .env to set your QUANTIZATION_BACKEND (fbgemm for Win, qnnpack for Linux)

2. Launch via Docker Compose

   docker-compose up -d

3. Access Services

   • Dashboard: http://localhost:7860
   • Inference Server: http://localhost:8000
   • Jupyter Lab: http://localhost:8888
   • TensorBoard: http://localhost:6006

---

📂 Data Preparation

The platform expects audio files in the following structure:

• data/raw/positive/: Put your wakeword audio files here (.wav, .flac, .mp3).
• data/raw/negative/: Put background noise and non-wakeword speech here.

The system will automatically create these directories on first run.

---

🏗️ Distributed Cascade Architecture

Production-Ready 3-Stage Pipeline for real-time wakeword detection:

⚡ Stage	🎯 Purpose	🧠 Model	📊 Metrics
Sentry (Edge)	Always-On Detection	MobileNetV3 + QAT	<1% FNR, <0.1% Energy
Judge (Local)	False Positive Filtering	Wav2Vec 2.0	>99% Accuracy
Teacher (Cloud)	Knowledge Distillation	Teacher-Student	10x Faster Training

🔬 Advanced Features: CMVN, EMA, Mixed Precision, FAH Metrics
📖 Architecture Deep Dive

---

What's New in v4.0

• 📉 New: Focal Loss implementation for superior hard-negative handling
• ⚡ New: QAT Accuracy Recovery pipeline (FP32 baseline to INT8 fine-tuning)
• 📏 New: Model Size Insight & Platform Constraints validation for Edge deployment
• ✨ New: Advanced GPU acceleration with Mixed Precision training
• 🚀 New: Comprehensive HPO (Hyperparameter Optimization) system
• 📦 New: Production-ready ONNX and TFLite export
• 🎯 New: Knowledge Distillation for 10x faster edge deployment
• 🔧 New: Raspberry Pi optimized models and configs

📄 License

MIT License - See LICENSE file for details

---

🚀 Happy Training! ⭐ Star us on GitHub!