Timecourse of phonological competition in causal and non-causal transformer models of speech processing

<p dir="ltr">Activation of targets (e.g., beaker), cohorts (e.g., beetle), rhymes (e.g., speaker), and unrelated words (e.g., carriage) in causal (sequential processing, only access to current and past speech signal) and non-causal (sequential processing but with access to past, current, and future speech signal) transformer models of speech processing. The causal model exhibits the primary characteristics of human phonological competition: activations are driven by current signal information, cohorts and rhymes have higher activation than unrelated words, and cohorts activate and peak earlier than rhymes.

The non-causal model exhibits a timecourse very dissimilar to that of humans. Replotted from Peng, L. et al., Interspeech, 2026. </p>