Cargando…

Dataset size considerations for robust acoustic and phonetic speech encoding models in EEG

In many experiments that investigate auditory and speech processing in the brain using electroencephalography (EEG), the experimental paradigm is often lengthy and tedious. Typically, the experimenter errs on the side of including more data, more trials, and therefore conducting a longer task to ens...

Descripción completa

Detalles Bibliográficos
Autores principales: Desai, Maansi, Field, Alyssa M., Hamilton, Liberty S.
Formato: Online Artículo Texto
Lenguaje:English
Publicado: Frontiers Media S.A. 2023
Materias:
Acceso en línea:https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9895838/
https://www.ncbi.nlm.nih.gov/pubmed/36741776
http://dx.doi.org/10.3389/fnhum.2022.1001171
_version_ 1784881931434852352
author Desai, Maansi
Field, Alyssa M.
Hamilton, Liberty S.
author_facet Desai, Maansi
Field, Alyssa M.
Hamilton, Liberty S.
author_sort Desai, Maansi
collection PubMed
description In many experiments that investigate auditory and speech processing in the brain using electroencephalography (EEG), the experimental paradigm is often lengthy and tedious. Typically, the experimenter errs on the side of including more data, more trials, and therefore conducting a longer task to ensure that the data are robust and effects are measurable. Recent studies used naturalistic stimuli to investigate the brain's response to individual or a combination of multiple speech features using system identification techniques, such as multivariate temporal receptive field (mTRF) analyses. The neural data collected from such experiments must be divided into a training set and a test set to fit and validate the mTRF weights. While a good strategy is clearly to collect as much data as is feasible, it is unclear how much data are needed to achieve stable results. Furthermore, it is unclear whether the specific stimulus used for mTRF fitting and the choice of feature representation affects how much data would be required for robust and generalizable results. Here, we used previously collected EEG data from our lab using sentence stimuli and movie stimuli as well as EEG data from an open-source dataset using audiobook stimuli to better understand how much data needs to be collected for naturalistic speech experiments measuring acoustic and phonetic tuning. We found that the EEG receptive field structure tested here stabilizes after collecting a training dataset of approximately 200 s of TIMIT sentences, around 600 s of movie trailers training set data, and approximately 460 s of audiobook training set data. Thus, we provide suggestions on the minimum amount of data that would be necessary for fitting mTRFs from naturalistic listening data. Our findings are motivated by highly practical concerns when working with children, patient populations, or others who may not tolerate long study sessions. These findings will aid future researchers who wish to study naturalistic speech processing in healthy and clinical populations while minimizing participant fatigue and retaining signal quality.
format Online
Article
Text
id pubmed-9895838
institution National Center for Biotechnology Information
language English
publishDate 2023
publisher Frontiers Media S.A.
record_format MEDLINE/PubMed
spelling pubmed-98958382023-02-04 Dataset size considerations for robust acoustic and phonetic speech encoding models in EEG Desai, Maansi Field, Alyssa M. Hamilton, Liberty S. Front Hum Neurosci Human Neuroscience In many experiments that investigate auditory and speech processing in the brain using electroencephalography (EEG), the experimental paradigm is often lengthy and tedious. Typically, the experimenter errs on the side of including more data, more trials, and therefore conducting a longer task to ensure that the data are robust and effects are measurable. Recent studies used naturalistic stimuli to investigate the brain's response to individual or a combination of multiple speech features using system identification techniques, such as multivariate temporal receptive field (mTRF) analyses. The neural data collected from such experiments must be divided into a training set and a test set to fit and validate the mTRF weights. While a good strategy is clearly to collect as much data as is feasible, it is unclear how much data are needed to achieve stable results. Furthermore, it is unclear whether the specific stimulus used for mTRF fitting and the choice of feature representation affects how much data would be required for robust and generalizable results. Here, we used previously collected EEG data from our lab using sentence stimuli and movie stimuli as well as EEG data from an open-source dataset using audiobook stimuli to better understand how much data needs to be collected for naturalistic speech experiments measuring acoustic and phonetic tuning. We found that the EEG receptive field structure tested here stabilizes after collecting a training dataset of approximately 200 s of TIMIT sentences, around 600 s of movie trailers training set data, and approximately 460 s of audiobook training set data. Thus, we provide suggestions on the minimum amount of data that would be necessary for fitting mTRFs from naturalistic listening data. Our findings are motivated by highly practical concerns when working with children, patient populations, or others who may not tolerate long study sessions. These findings will aid future researchers who wish to study naturalistic speech processing in healthy and clinical populations while minimizing participant fatigue and retaining signal quality. Frontiers Media S.A. 2023-01-20 /pmc/articles/PMC9895838/ /pubmed/36741776 http://dx.doi.org/10.3389/fnhum.2022.1001171 Text en Copyright © 2023 Desai, Field and Hamilton. https://creativecommons.org/licenses/by/4.0/This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
spellingShingle Human Neuroscience
Desai, Maansi
Field, Alyssa M.
Hamilton, Liberty S.
Dataset size considerations for robust acoustic and phonetic speech encoding models in EEG
title Dataset size considerations for robust acoustic and phonetic speech encoding models in EEG
title_full Dataset size considerations for robust acoustic and phonetic speech encoding models in EEG
title_fullStr Dataset size considerations for robust acoustic and phonetic speech encoding models in EEG
title_full_unstemmed Dataset size considerations for robust acoustic and phonetic speech encoding models in EEG
title_short Dataset size considerations for robust acoustic and phonetic speech encoding models in EEG
title_sort dataset size considerations for robust acoustic and phonetic speech encoding models in eeg
topic Human Neuroscience
url https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9895838/
https://www.ncbi.nlm.nih.gov/pubmed/36741776
http://dx.doi.org/10.3389/fnhum.2022.1001171
work_keys_str_mv AT desaimaansi datasetsizeconsiderationsforrobustacousticandphoneticspeechencodingmodelsineeg
AT fieldalyssam datasetsizeconsiderationsforrobustacousticandphoneticspeechencodingmodelsineeg
AT hamiltonlibertys datasetsizeconsiderationsforrobustacousticandphoneticspeechencodingmodelsineeg