DL captures human speech perception both *qualitatively* & *quantitatively* (R2>96%) for over 400 combinations of exposure and test items. Yet, previous DL models fail to capture important limitations. Specifically, we find that DL seems to proceed by remixing prev experience 2/2.

Very excited about this: putting distributional learning (DL) models of adaptive speech perception to a strong, informative test by Maryann Tan. We use Bayesian ideal observers & adapters to assess whether DL predicts rapid changes in speech perception 1/2.

Join millions who have switched to Grok.

Great opportunity!

RT @SimonJGreenhill: Do you want to know how much it costs to publish open access in a journal? I got sick of looking this up, so made th….

RT @eegmanylabs: 🚨Exciting news! We now have the first-ever complete #EEGManyLabs replication. This large-scale multi-site study revisits a….

RT @maxplanckpress: More than 40 percent of #postdocs leave academia. Those who landed a coveted faculty position were more likely to have….

RT @UoR_BrainCogSci: Please share: Postdoctoral position available in high-resolution retinal imaging and video eye-tracking at the Univers….

Feedback welcome! & see @_wbushong_ 's thesis for more on this topic, and her very cool computational simulations--presenting a new effort to better understand what can be inferred from the types of data collected in studies on subcat infomaintenance during speech perception. /n.

But Bayesian mixed-effect analyses also identify a previously undocumented tendency in all 4 datasets--unexpected under all existing accounts. We present initial simulations that suggests that a combination of attentional lapses & ideal integration might explain the data. /3.

We show existing evidence is compatible with ideal maintenance & integration of uncertainty, and derive a stronger test of that hypothesis. 2 re-analysis & 2 new experiments find that the ideal observer's predictions fit listeners' behavior better than previous proposals /2.

We revisit the classic work by Connine and colleagues, and show why its results are often misinterpreted. Using an ideal observer framework, we derive what would be expected if listeners maintained and integrated subcategorical information beyond word boundaries. /1.

For how long is information about past speech input available in short-term memory? What evidence would inform this question? What does (not) follow from previous work on this question? Delighted to see this work by @klintonbicknell @_wbushong_ in JML

RT @UoR_BrainCogSci: A new study from co-author Chigusa Kurumada reveals that difficulties in adapting to changes in speech patterns may af….

Great, super accessible, 3-step tutorial for running experiments online by Chigusa Kurumada ( @UoR_BrainCogSci @UofR.

Can't wait for this to arrive at my door! (t - 1.5 months).

A wonderful resource for anyone working with R & Praat to analyze phonetic & speech data.

Another neat paper by Xin Xie and Chigusa Kurumada: an initial attempt to measure adaptive changes with prolonged, repeated exposure to second language-accented speech (over 4 weeks) + insightful discussion of methodological challenges that arise

RT @ShravanVasishth: For the Bayes-curious: There are still some places open for my one-week long introduction to Bayesian statistics (spec….

RT @klintonbicknell: Duolingo AI is hiring at many levels: New PhDs, Senior AI Research Engineers, and an AI Research Manager! Come help us….

I think the results conceptually replicate for vowel categorization what @BozenaPajak found for durational lengthening Different cues, different contrasts, different language backgrounds, with results pointing to the same mechanism. Nice!.