Neural network word/false-alarm discriminators for improved keyword spotting

Abstract

Two neural networks which are trained on examples of whole words to perform the task of keyword spotting are described. One network is a temporally constrained self-organizing feature map. The other network is a time-delay network which has been modified by the addition of recurrent connections. These networks were tested as secondary processors to a conventional (non-neural network) wordspotter. In this scenario, the conventional system screened incoming speech for potential keywords which are passed to the networks for the final accept/reject determination. The database used for testing contains 20 key words. The test results are summarized in receiver operator characteristic (ROC) curves. These initial results indicate that wordspotting performance is helped by the application of neural network word discriminators as secondary processors. The percentage of keywords recognized was improved at all false alarm rates.< >