Flexible Stand-Alone Keyword Recognition Application Using Dynamic Time Warping

We introduce a Query-by-Example (QbE) application for smart-phone devices that implements a recently proposed memory-efficient dynamic programming algorithm [1] for the task of keyword search. The application compares acoustic keywords with the audio input from the microphone and reacts to detected keywords with actions in the phone. These keywords are recorded by the user, who also defines what actions will be performed by each one. One of these keywords is defined to be a trigger keyword, which is used to wake up the system and thus reduce false detections. All keywords can be freely chosen by the user. In Monitor mode, the application stays listening to audio acquired through the microphone and reacts when the trigger + some keyword are matched. All processing is done locally on the phone, which is able to react in realtime to incoming keywords. In this paper we describe the application, review the matching algorithm we used and show experimentally that it successfully reacts to voice commands in a variety of acoustic conditions.