A Sparse Neural Code for Some Speech Sounds but Not for Others

Mathias Scharinger,Alexandra Bendixen,Nelson J Trujillo-Barreto,Jonas Obleser

doi:10.1371/journal.pone.0040953

Abstract

The precise neural mechanisms underlying speech sound representations are still a matter of debate. Proponents of ‘sparse representations’ assume that on the level of speech sounds, only contrastive or otherwise not predictable information is stored in long-term memory. Here, in a passive oddball paradigm, we challenge the neural foundations of such a ‘sparse’ representation; we use words that differ only in their penultimate consonant (“coronal” [t] vs. “dorsal” [k] place of articulation) and for example distinguish between the German nouns Latz ([lats]; bib) and Lachs ([laks]; salmon). Changes from standard [t] to deviant [k] and vice versa elicited a discernible Mismatch Negativity (MMN) response. Crucially, however, the MMN for the deviant [lats] was stronger than the MMN for the deviant [laks]. Source localization showed this difference to be due to enhanced brain activity in right superior temporal cortex. These findings reflect a difference in phonological ‘sparsity’: Coronal [t] segments, but not dorsal [k] segments, are based on more sparse representations and elicit less specific neural predictions; sensory deviations from this prediction are more readily ‘tolerated’ and accordingly trigger weaker MMNs. The results foster the neurocomputational reality of ‘representationally sparse’ models of speech perception that are compatible with more general predictive mechanisms in auditory perception.

Highlights

Recent decades of research in auditory neuroscience have led to a better understanding of cortical mechanisms subserving the perception of single speech sounds, of combinations of speech sounds, and of whole words (e.g. [1])
Amongst various linguistically-based approaches, some consider cognitive speech sound representations to be entirely faithful to the physical signal and stipulate very detailed and finegrained units (e.g. Exemplars, [5,6,7,8,9]), sometimes linked to episodic memory [10]
One participant had to be excluded due to a hearing ailment, one participant due to technical problems (EEG data of one block were not recorded), and one additional participant due to low signal-tonoise ratio in the EEG. This resulted in a total of seventeen participants (41% females, mean age 25.4, SD 2.2) whose data were used for the event-related potential (ERP) analyses

Summary

Introduction

Recent decades of research in auditory neuroscience have led to a better understanding of cortical mechanisms subserving the perception of single speech sounds, of combinations of speech sounds, and of whole words (e.g. [1]). Amongst various linguistically-based approaches, some consider cognitive speech sound representations to be entirely faithful to the physical signal and stipulate very detailed and finegrained units (e.g. Exemplars, [5,6,7,8,9]), sometimes linked to episodic memory [10]. If such models held, flexibility of speech sound representations would have to be achieved through learning of many exemplars in many different situations. These sounds are considered to be based on a more precise, rigid neural code

Methods

Results

Conclusion