Comparing pre-linguistic normalization models against US English listeners’ vowel perception

Abstract

One of the central computational challenges for speech perception is that talkers differ in pronunciation--i.e., how they map linguistic categories and meanings onto the acoustic signal. Yet, listeners typically overcome these difficulties within minutes (Clarke & Garrett, 2004; Xie et al., 2018). The mechanisms that underlie these adaptive abilities remain unclear. One influential hypothesis holds that listeners achieve robust speech perception across talkers through low-level pre-linguistic normalization. We investigate the role of normalization in the perception of L1-US English vowels. We train ideal observers (IOs) on unnormalized or normalized acoustic cues using a phonetic database of 8 /h-VOWEL-d/ words of US English (N = 1240 recordings from 16 talkers, Xie & Jaeger, 2020). All IOs had 0 DFs in predicting perception—i.e., their predictions are completely determined by pronunciation statistics. We compare the IOs’ predictions against L1-US English listeners’ 8-way categorization responses for /h-VOWEL-d/ words in a web-based experiment. We find that (1) pre-linguistic normalization substantially improves the fit to human responses from 74% to 90% of best-possible performance (chance = 12.5%); (2) the best-performing normalization accounts centered and/or scaled formants by talker; and (3) general purpose normalization (C-CuRE, McMurray & Jongman, 2011) performed as well as vowel-specific normalization.