MULTI-TASK LEARNING FOR PERSONALIZED KEYWORD SPOTTING: abstract simplified (18153932)

Systems and techniques are described for processing audio data, specifically for personalized keyword spotting through multi-task learning (PK-MTL). The process involves obtaining an audio sample and generating representations of both a keyword and a speaker based on the sample. These representations are then used to determine a similarity score against a reference representation, which is associated with the keyword and/or the speaker. Based on this similarity score and a threshold, a keyword spotting (KWS) output is generated to determine if the audio sample includes the target keyword.