Speech recognition can be modeled as matching test signals against noise-masked talker templates

Evidence

A template-matching procedure was applied to predict the responses obtained in experiment 1. Two talker-specific template sets were considered, consisting of all speech tokens from each talker (i.e., three templates for each CV). The templates were mixed with random white noise at the test-signal SNR and compared to the experimental stimuli. The distances between the models’ internal representations of the test signals and the templates were obtained using a standard dynamic time warping (DTW) algorithm (Sakoe and Chiba 1978 ). The template-matching procedure was conducted nine times with newly generated random noise for the templates. The “correct” template always contained the same speech token as the test signal, while the masking noise differed. In each run, the template showing the smallest distance to the test signal was selected. The modeled responses were converted to proportions of responses. — Dijk, Pim van, Physiology, Psychoacoustics and…, 5606

Categorization

• XREF: Template-matching with DTW connects to exemplar-based speech perception theory; the repeated-noise-injection averaging parallels psychophysical detection experiments.