SYSTEM AND METHOD FOR ON-DEVICE OPEN-VOCABULARY KEYWORD SPOTTING

摘要

Device (1) comprising a keyword encoder (2), an acoustic encoder (3), a keyword detector (4); - wherein the keyword encoder (2) is a first neural network configured to determine, for each keyword of a set of keywords (abc, bac, cba, cab), at least one parameter, the at least one parameter being determined based on a keyword representation (7), said keyword representation (7) being provided as input to the keyword encoder (2), the keyword encoder (2) being configured to communicate the at least one parameter to the keyword detector (4), for each keyword of the set of keywords (abc, bac, cba, cab); - wherein the acoustic encoder (3) consists in a second neural network trained for automatic speech recognition, the acoustic encoder (3) being configured to take as input a set of feature vectors, each feature vector of the set of feature vector comprising at least one feature characterizing an acoustic signal (5), the acoustic encoder (3) being further configured to provide as output a set of intermediate feature vectors (6), each intermediate feature vector of the set of intermediate feature vectors (6) comprising at least one feature, and being input to the keyword detector (4); - wherein the keyword detector (4) is a third neural network configured, by the at least one parameter, to predict, for each keyword in the set of keywords (abc, bac, cba, cab), a probability of detection of said keyword in the acoustic signal, the probability of detection (P1,P2,P3,P4) of said keyword being based on the set of intermediate feature vectors (6) provided by the acoustic encoder (3).