A complete voice-driven experience in applications such as wearable electronics requires always-on keyword monitoring, which is prohibitively power consuming using current speech recognition methods. In this work, we propose an ultra-low power voice command recognition system that is designed to recognize short commands such as 'Hi Galaxy'. To achieve power-efficient designs, the system uses adaptive feature pre-selection such that only a subset of all available features are selected and extracted based on the noise spectrum. The backend classifier, supporting adaptive feature selection, is enabled by a novel multi-band deep neural networks (DNNs) model that processes only the selected features at each decision. In experiments, our adaptive scheme achieves comparable accuracy and improved efficiency using an average of 5 spectral feature bands, than a generic fully-connected DNNs model using the full speech spectrum. The system makes a recognition decision every 40ms on 1.2s of buffered speech and consumes ～ 230μW of power, thus promising low-power, low-complexity and robust application-specific voice recognition.
展开▼
