The LKW project is aimed at designing low-power algorithms and systems for admission control to speech systems: i.e., detecting foreground speech, recognizing leading keywords and verifying speakers on a continuously-on wearable device. Our goal is to consume under 10 mW average on generic embedded hardware available today and under 100uW on custom hardware.
LKW is predicated on the simple insight that admission control for speech is usually simpler than speech recognition. In other words, it should be simpler to detect whether an authorized person is currently speaking to a device than to tell what they are speaking. When a wearable device is designed as an accessory to a mobile device such as a phone, this suggests a natural division of responsibility: the wearable, which usually has 10-20x smaller battery capacity and 100-1000x less memory than the phone, performs admission control. It wakes up the phone when appropriate, and yields to the phone to recognize speech. This decoupled approach stands in contrast to current commercial devices that seek to perform full speech recognition at ultra low power at the cost of limiting recognized speech and requiring full custom ASIC implementations.
Our initial results are promising: we can recognize leading keywords and verify speakers at well over 90% true positive rates (at below 5-10% false positives) using 10-100x fewer instructions than commercial speech-recognition based pipelines. Our code is currently available via the LKW repository. We are sharing with select early collaborators: contact us if you would like access.