LAUSR

SUBMISSION DETAILS Presentation Type Either Poster or Oral Presentation Presentation Abstract Summary Recent advances in machine learning raised the prospect of developing quantitative models of human perception that can handle realistic sensory input. In this paper, we focus on the question of phonetic category perception, i.e. the way we perceive basic speech sounds (roughly consonants and vowels), which is largely determined by the language(s) to which we were exposed as a child. For example, native speakers of Japanese have a hard time discriminating between American English /r/ and /l/, a phonetic contrast that has no equivalent in Japanese. We show that typical GMM-HMM Automatic Speech Recognition (ASR) systems trained on large corpora of continuous speech correctly predict several perceptual effects observed in humans. Our work illustrates the value of considering large-scale machine learning systems in the context of modeling human perception. Paper Upload (PDF) CCN2017-3.pdf Co-author Information