NanoWakeWord is an advanced, lightweight wake word detection engine designed to create custom, high-accuracy models effortlessly. Its unique, adaptive framework optimizes training, making it easy to tailor solutions for various applications while ensuring high performance.

Key Features

• Custom Model Architectures: Choose from a variety of neural network architectures optimized for different scenarios. Select from options such as DNN, RNN, CNN, LSTM, GRU, and more, according to your specific needs.

Automated ML Engineering for Peak Performance

NanoWakeWord's core lies in its intelligent data-driven configuration engine that tailors training for optimal results:

• Adaptive Architectural Scaling: Automatically adjusts the model complexity based on dataset characteristics.
• Optimized Training Strategy: Implements multi-stage learning rate schedules, ensuring efficient training.
• Hardware-Aware Performance Tuning: Profiles hardware to maximize efficiency and throughput.
• Automatic Pre-processing: Handles audio format standardization seamlessly.

Comprehensive Data Pipeline

The engine includes a robust data pipeline that handles the lifecycle from raw audio to optimized features. Key elements include:

• Phonetic Adversarial Negative Generation, which creates counter-examples to reduce false positives.
• Dynamic Augmentation, adding varied acoustic conditions for comprehensive training.
• Memory-Mapped Data Handling, enabling the management of large datasets effortlessly.

State-of-the-Art Optimization

The training process incorporates modern techniques to ensure model reliability and robustness:

• Checkpoint Ensembling for enhanced accuracy by averaging the weights of stable models.
• Transparent Live Dashboard, providing real-time visibility into training parameters and performance.

Deployment-Optimized Inference Engine

Designed for efficiency, the inference engine allows for seamless deployment across different environments, from edge devices to powerful servers. Features include:

• Stateful Streaming Architecture for low-latency predictions.
• Universal Model Export in industry-standard formats like ONNX.
• Integrated On-Device Processing, utilizing Voice Activity Detection and Noise Reduction for optimal performance in real-world applications.

Getting Started

Begin by installing NanoWakeWord using pip:

pip install nanowakeword

To train custom models, install the full package:

pip install "nanowakeword[train]"

Define your project using a .yaml configuration file that manages data paths and pipeline stages, ensuring repeatability and clarity.

Conclusion

NanoWakeWord stands out as a powerful, customizable solution for wake word detection, empowering developers to create high-performance applications with ease. With a strong emphasis on automation, efficiency, and user-friendly processes, it simplifies the development of advanced voice recognition technology.