BCS 152 Lecture Notes - Lecture 13: Labeled Data, Neural Adaptation, Causal Inference
Document Summary
Result: accept more sounds as sh even if s-like. Original category boundary when not exposed to speaker w/ shifts in articulation. Accept more sounds as s even if sh-like. Given more variability in the acoustic cues, the listener takes longer to adapt but what they learn is more generalizable to new sounds in the same category or to new speakers situation. Listeners are flexible- we don"t use the same category boundary for every. Listeners can adapt to shifted production of a sound contrast even when they. People infer categories of new speakers in both supervised and unsupervised. Adaptation generalizes to new sound contrasts along similar acoustic don"t get labeled data contexts dimensions. Expose to full distribution of a speaker"s productions. Adaptation to one sound contrast (/b/-/p/) generalizes to other sounds that. Adaptation to accents & dialects contrast along the same feature (/d/-/t/) Clarke & garrett : listen to sentence, and presented word appears.