Ant-Brain/EfficientWord-Net

Implements Siamese Network architecture (inspired by FaceNet) with TensorFlow Lite for efficient real-time inference, requiring only 3-4 user-recorded samples to generate custom hotword embeddings. Supports both single and multiple concurrent hotword detection via sliding-window audio processing with configurable confidence thresholds and relaxation timing to minimize false positives. Includes pre-trained models (Resnet_50_Arc_loss, First_Iteration_Siamese) and out-of-the-box reference files for common wakewords (Mycroft, Google, Alexa, etc.), with optional C++ inference support.

301 stars.