Figure 1.

A flowchart representation of the keyword-spotting signal processing pipeline. Red arrows indicate flow of data through the pipeline. Dotted lines with circles indicate models that were trained on the training dataset are used in this step for both the training and testing data. Training is performed using a visually presented keyword reading paradigm, and testing occurs across an auditory keyword repetition task. This study implements a two-stage detector; one neural VAD template is correlated across the testing dataset, and peak-picking indicates a detected utterance. When an utterance is detected, a discriminative classifier is used to decide if the utterance was a keyword or non-keyword speech. Channel downselection, normalization parameters, neural templates, feature dimensionality reduction, and classifiers are all trained on the reading (training) task and applied to the repetition (testing) task to simulate how keyword spotting would realistically perform in a separate recording session.