Wake-Word Detection Using Audio Embeddings in Noisy Environments

The patent introduces acoustic word embeddings as an intermediary layer between the audio input and the wake-up command detection. The embedding model transforms raw audio signals into semantic representations that capture the meaning of wake-up phrases while being invariant to environmental noise, cross-talks, and acoustic variations. This intermediary representation allows the system to achieve high detection accuracy without being directly affected by noisy environmental factors.

The patent replaces traditional mechanical keyword matching systems with a machine learning-based embedding model. Instead of using rigid pattern matching that is sensitive to noise, the system employs neural network-based acoustic word embeddings that learn robust semantic representations from data, enabling accurate wake-up detection in challenging acoustic environments.

The patent performs preliminary action by pre-training the acoustic word embedding model offline using large datasets. The complex computational work of learning robust audio representations is done in advance during model training, allowing the deployed system to use the pre-trained embeddings for efficient real-time wake-up detection without bearing the full computational burden of training.

The patent segments the wake-up detection system into distinct components: an acoustic word embedding model for feature extraction and a classification layer for decision-making. This segmentation allows each component to be optimized independently, with the embedding model handling the complex task of noise-robust feature extraction while the classifier remains relatively simple.