Voice-based user recognition

Abstract

Techniques for performing voice-based user recognition are described. When a device receives audio data corresponding to a spoken user input, the device may generate spoken user input embedding data representing speech characteristics of the spoken user input. The device may also identify user embedding data representing speech characteristics of a user known to the device. A machine learning (ML) model may process the spoken user input embedding data to generate reduced spoken user input embedding data including a reduced number of dimensions, where the reduced number of dimensions are functions of the higher number of dimensions of the spoken user input embedding data. Moreover, the ML model may process the user embedding data to generate reduced user embedding data including a reduced number of dimensions, where the reduced number of dimensions are functions of the higher number of dimensions of the user embedding data and are tuned to characteristics of users of the device (as opposed to a representation of all possible users). A comparison of the reduced spoken user input embedding data and the reduced user embedding data may be used to determine whether the user spoke the user input.