Dual learning systems in talker identification: the effects of language, accent, and feedback

Talker identification categorizes variable speech signals into stable talker representations, a process facilitated by language and accent familiarity. The dual learning systems (DLS) model posits that speech category learning involves a "reflective" system based on explicit rules and a "reflexive" system based on stimulus-reward associations, with reflexive learning dominating in later stages. In this study, we leverage the DLS framework to investigate talker learning by training Mandarin-speaking listeners to identify talkers in native (Mandarin) and nonnative languages with native (English) or nonnative, but familiar accents (Mandarin-accented English) contexts. Listeners received either using full (e.g., Incorrect. It's Talker 1) or minimally informative (e.g., Incorrect) feedback, encouraging reflective or reflexive learning, respectively. We assessed identification performance through accuracy and response times and analyzed the underlying decision processes using drift diffusion models. Results showed that language and accent familiarity improved accuracy and response times. At later training stages, minimal feedback, which promotes reflexive learning according to the DLS model, facilitated faster identification and more efficient decision-making, particularly in the nonnative language context (English). The findings highlight the benefit of reflexive learning in talker identification through improved response efficiency and the need to consider decision dynamics in this process. The data, materials, and analysis code are available online ( https://osf.io/g7r9q/ ).