Loading...
An open-source Python framework for building custom wake-word detection models, from audio data preparation and neural network training to model optimization and deployment.
A wake word is a specific word or phrase that activates a voice-controlled system. NanoWakeWord gives developers the tools to create detection models for their own custom wake words instead of depending on a fixed set of supported phrases.
It is a complete development framework that covers the process of creating a wake-word model. Developers can prepare audio datasets, choose neural network architectures, train and evaluate models, and export them for real-world applications.
NanoWakeWord supports both embedding-based and end-to-end approaches, providing flexibility in how wake-word detection systems are developed.
Everything needed to develop, train and run custom wake-word detection models in one framework.
Prepare training data using synthetic speech, phonetic adversarial negative generation and audio augmentation. Introduce background noise, different SNR levels, room impulse responses, pitch variations and filtering to create diverse training conditions.
Build wake-word detection models with a wide range of neural network architectures. Use lightweight networks for resource-constrained devices or experiment with more advanced architectures for different detection requirements.
Train models with loss-aware hard-example sampling, configurable training pipelines, memory-mapped datasets, checkpointing and weight averaging. Knowledge distillation also makes it possible to transfer learned behavior into smaller student models.
Evaluate trained models, monitor training performance and improve detection behavior. Analyze false positives and false negatives to understand how a model performs in practical environments.
Export trained models to ONNX, PyTorch and supported custom formats. Prepare models for integration into different applications and computing environments.
Run continuous audio detection with streaming inference. Optional VAD, noise reduction, debounce and patience mechanisms help manage predictions during real-time operation.
Different applications have different requirements. NanoWakeWord supports a variety of neural network architectures, from compact models to more advanced sequence-based networks. Custom architectures can also be integrated into the training workflow.
A straightforward development process helps you move from an initial wake-word dataset to a model ready for integration.
Collect or generate positive and negative audio samples for the target wake word.
Select a supported architecture or integrate a custom neural network.
Configure training, apply augmentation and use advanced sampling techniques.
Test detection behavior and refine the model based on evaluation results.
Export the trained model and integrate it into your application.
Wake-word detection can be used in many environments. NanoWakeWord provides model export and inference options to support different deployment needs.
Run detection directly on a local computer or device.
Integrate optimized models into resource-constrained hardware.
Run models through server-based inference with API access controls.
Combine local detection with remote verification when needed.
Explore the source code, install NanoWakeWord and start developing custom wake-word models for your applications.
Developed by Arcosoph
