SparrKULee: A Speech-Evoked Auditory Response Repository from KU Leuven, Containing the EEG of 85 Participants

Abstract

Researchers investigating the neural mechanisms underlying speech perception often employ electroencephalography (EEG) to record brain activity while participants listen to spoken language. The high temporal resolution of EEG enables the study of neural responses to fast and dynamic speech signals. Previous studies have successfully extracted speech characteristics from EEG data and, conversely, predicted EEG activity from speech features. Machine learning techniques are generally employed to construct encoding and decoding models, which necessitate a substantial quantity of data. We present SparrKULee, a Speech-evoked Auditory Repository of EEG data, measured at KU Leuven, comprising 64-channel EEG recordings from 85 young individuals with normal hearing, each of whom listened to 90–150 min of natural speech. This dataset is more extensive than any currently available dataset in terms of both the number of participants and the quantity of data per participant. It is suitable for training larger machine learning models. We evaluate the dataset using linear and state-of-the-art non-linear models in a speech encoding/decoding and match/mismatch paradigm, providing benchmark scores for future research.

1. Summary

In order to study the neural processing of speech, recent studies have presented natural running speech to participants while recording an electroencephalogram (EEG). Currently, regression is used to either decode features of the speech stimulus from the EEG (also known as a backward model) [1,2,3,4,5], to predict the EEG from the speech stimulus [1,6] (forward model), or to transform both EEG and speech stimulus to a shared space [7,8] (hybrid model). Deep neural networks have recently been proposed for auditory decoding and have obtained promising results [4,5,9,10,11,12].

All previously mentioned methods require EEG recordings of the participants with strict time alignment to the speech stimulus. This time alignment is necessary due to the time-locked neural tracking of the speech stimulus at a millisecond scale (e.g., auditory brainstem responses (ABRs)), which can last up to 600 ms [13]. As these data are personal and expensive to collect, there is a need for more public datasets that researchers can use to benchmark and train their models.

Table 1. Overview of currently publicly available single-speaker datasets.

Dataset	Ref	Speech Material	Language	Participants	Average Time per Participant (min)	Total Time (min)
Broderick	[25]	Clean speech	English	19	60	1140
		Time-reversed speech		10	60	600
		Speech-in-noise		21	30	630
DTU Fuglsang	[26]	Clean speech	Danish	18	8.3	150
Etard	[27]	Clean speech	English	18	10	180
		Speech-in-noise		18	30	540
		Foreign language speech	Dutch	12	40	480
Weissbart	[28]	Clean speech	English	13	40	520
Brennan	[29]	Clean speech	English	49	12.4	610
Vanheuseden	[30]	Clean speech	English	17	24	410
SparrKULee		Clean speech	Dutch	85	110	9320
		Speech-in-noise		26	28.5	740

presents an overview of currently available public datasets of EEG recordings of people listening to natural speech. These studies have generated 87.7 h of EEG data from 133 participants listening to clean speech and speech-in-noise in their native language. However, this quantity of data is relatively small compared to datasets in other domains, such as automatic speech recognition, and needs to be increased for training models due to the low signal-to-noise ratio of auditory EEG. Additionally, combining the data from these studies for model training is challenging due to differences in the authors’ signal acquisition equipment, measurement protocols, and preprocessing methods. SparrKULee provides a large homogeneous dataset to train and evaluate models effectively, alleviating the need for combining multiple heterogeneous datasets. Its utilization has been facilitated through two organized challenges [14,15,16], which have catalyzed the dataset’s adoption, leading to a range of advancements and novel model architectures [17,18,19,20,21,22,23,24].

2. Methods

For our dataset (SparrKULee), we conducted an EEG experiment in which 85 participants were recruited and presented with speech stimuli for a duration ranging between 90 and 150 min, divided into 6 to 10 recordings (i.e., an uninterrupted period in which a participant listens to a stimulus), totaling 168 h of EEG data. To validate the obtained dataset, we employed state-of-the-art linear [2,8,31] and deep learning models [12], in participant-specific and participant-independent training scenarios (note that the cited studies used different, smaller datasets). These models can serve as benchmarks for comparison in future research. Our dataset is publicly available on the RDR KU Leuven website https://doi.org/10.48804/K3VSND (accessed on 1 June 2024).

We defined a trial as an uninterrupted recording lasting around 15 min. We defined a session as the complete set of trials and pre-screening activities that a participant underwent from the moment they entered the room until the moment they left. Stimulus, in our study, referred to the speech audio files that we presented to the participants during the experiment, which were designed to elicit specific responses from their brains. Figure 1 provides a high-level overview of the different parts of a session.

2.1. Participants

Between October 2018 and September 2022, data were collected from 85 participants (74 females/11 males, 21.4 ± 1.9 years (sd)). Inclusion criteria for this study were young (18–30 years), normal-hearing adults (all hearing thresholds ≤ 30 dB HL, for 125–8000 Hz), with Dutch/Flemish as their native language. Before commencing the EEG experiments, participants read and signed an informed consent form approved by the Medical Ethics Committee UZ KU Leuven/Research (KU Leuven, Belgium) with reference S57102. All participants in this dataset explicitly consented to share their pseudonymized data in a publicly accessible dataset. This dataset is a subset of our larger proprietary dataset containing data from participants who did not give consent to share their data. Additionally, the participants completed a questionnaire requesting general demographic information (age, sex, education level, handedness [32]) and diagnoses of hearing loss and neurological pathologies. Participants indicating any neurological or hearing-related diagnosis were excluded from the study. Last, the medical history and the presence of learning disabilities were questioned as research has shown that serious concussions, the medication used to treat, for example, insomnia [33] and learning disabilities such as dyslexia can affect brain responses [34,35]. Therefore, that information was used to screen out participants with possibly divergent brain responses.

2.2. Experimental Procedure

In this section the experimental procedure for the behavioral screening and subsequent EEG measurement are explained.

2.2.1. Behavioral

First, we measured the air conduction thresholds using the Hughson-Westlake method [36] for frequencies from 125 to 8000 Hz (see Figure 2). Participants with hearing thresholds > 30 dB HL were excluded.

Secondly, we used the Flemish matrix test [37] to determine each participant’s speech reception threshold (SRT, the signal-to-noise ratio (SNR) at which 50% speech understanding is achieved). The test consisted of 3 lists (2 for training, 1 for evaluation) of 20 sentences following the adaptive procedure of Brand et al. [38]. Each sentence had a fixed syntactic structure of 5 words: name, verb, numeral, color, and object [e.g., “Lucas telt vijf gele sokken” (“Lucas counts five yellow socks”)]. After each sentence, participants were asked to indicate the heard sentence using a 5 × 11 matrix containing ten possibilities for each word and a blank option. The order of the three lists was randomized across participants. The last SNR value was used as an estimate of the SRT. The lists were presented to the participants using electromagnetically shielded Ethymotic ER-3A insert phones, binaurally at 62 dBA for each ear. Luts et al. [37] presented the list to the participants monoaurally to the best ear and obtained an average SRT of −8.7 dB SNR when using the results of the third list of the adaptive procedure. During the first repetitions, they reported a significant training effect, which disappeared starting from the third repetition. In our setup, binaural stimulation was chosen to be close to our EEG data acquisition setup. Figure 3 shows the histogram of the obtained SRT over participants in our study. Participants scored an average value of −8.9 dB ± 0.6 (sd), similar to results obtained by Luts et al. [37].

2.2.2. EEG

All participants listened to 6, 7, 8, or 10 trials, each of approximately 15 min. The order of all the trials was randomized per participant. After each trial, a question about the stimulus content was asked to determine attention to and comprehension of all audio stimuli. As the questions were not calibrated, they merely motivated the participant to pay attention to the stimulus, and no further analysis of the answers is provided. After three trials, the participants were asked if they wanted to have a short break.

Table 2. Overview of the experimental procedure.

Experimental Procedure	Required Time (min)	Cumulative Time (min)
Fill in informed consent	5	5
Fill in questionnaire	5	10
Pure tone audiometry	15	25
Speech audiometry (matrix test)	25	50
Fit EEG equipment	15	65
Listen to 3 stimuli	50	115
First break	5	120
Listen to 3 stimuli	50	170
Second break	5	175
Krios scan of EEG electrode positions	10	185
Listen to 3 stimuli	50	245

shows an overview of the experiment and timing.

We used different categories of stimuli:

• Reference audiobook to which all participants listened, made for children and narrated by a male speaker. The length of the audiobook was around 15 min.
• Audiobooks made for children or adults. To keep the trial length around 15 min, some audiobooks were split into different partswhen the length exceeded 15 min.
• Audiobooks with noise made for children, to which speech-weighted noise was added, as explained below, to obtain an SNR of 5 dB.
• Podcasts from the series “Universiteit van Vlaanderen” (University of Flanders) [39]. Each episode of this podcast answers a scientific question, lasts around 15 min, and is narrated by a single speaker.
• Podcasts with video from the series “Universiteit van Vlaanderen” (University of Flanders) [39], while video material of the speaker was shown. The video material can be found on the website of Universiteit van Vlaanderen for each podcast separately. The video contains parts where the face of the speaker is visible.

• The dynamic range of the podcasts and podcasts with video was compressed by the producers of the stimuli, while that of the audiobooks was not.

The dataset collection consisted of two main session types, ses-shortstories01 and ses-varyingstories, differing in the presented stimuli. Each participant undertook one session. An overview of the experiment and timing can be found in Table 2, while Figure 4 summarizes which stimuli were used for each participant in each session.

2.2.3. Ses-Shortstories01

For this session type, data from 26 participants are available. It includes ten different parts of audiobooks for children. Two audiobooks, audiobook_1 and audiobook_4, were narrated by male speakers, the other by female speakers. audiobook_3, audiobook_5, and audiobook_6 were narrated by the same speaker. Two out of ten trials were randomly chosen for each participant and presented in speech-weighted noise (SNR = 5 dB). Additionally, 3 subjects listened to a different version of audiobook_1. For that experiment, the audiobook was cut in 2 halves (audiobook_1_1, audiobook_1_2, respectively), and a pitch-shifted version was created and used for each half (audiobook_1_1_shifted, audiobook_1_2_shifted, respectively). More information about the pitch shifting and additional experiments can be found in the work of Algoet et al. [40]. Finally, there was one control condition in which the first 5 min of audiobook_1 was presented to a subject who had no insert phones inserted (audiobook_1_artefact).

2.2.4. Ses-Varyingstories

For the ses-varyingstories type, data from 59 participants are available. Ses-varyingstories had a fixed reference audiobook_1 (which was presented to all subjects), an audiobook of around 30 min split into two parts, and three to five different podcasts per participant, chosen to keep an even distribution of the sex of the speaker. The stimuli were changed every 2 to 8 participants.

2.3. Data Acquisition

2.3.1. EEG

All recording sessions were conducted at the research group ExpORL of KU Leuven, in a triple-walled, soundproof booth equipped with a Faraday cage to reduce external electromagnetic interference. Participants were instructed to listen to the speech while seated and minimize muscle movements. They were seated in a comfortable chair in the middle of the booth.

We recorded EEG using a BioSemi ActiveTwo system with 64 active Ag-AgCl electrodes and two additional electrodes for the common electrode (CMS) and current return path (DRL). In addition, two mastoid electrodes and the BioSemi head caps were used, containing electrode holders placed according to the 10-20 electrode system.

To ensure proper electrode placement for each participant, we first measured their head size (from nasion to inion to nasion) and selected an appropriate cap. Mastoid locations were scrubbed with Nuprep and cleaned with alcohol gel. The mastoid electrodes were then attached using stickers and held with tape.

The electrode cap was placed on the participant’s head from back to front, with ears placed through gaps in the cap. The closing tape at the bottom was secured, and a visual assessment was performed to ensure proper fit. The cap was adjusted so that the distances between the nasion and the electrode Cz and the inion and the electrode Cz were equal, as well as the distance between the left and right ears and the Cz electrode. Electrode gel was applied to the cap holes, and the electrodes were placed gently. The battery, electrode cables, and mastoid electrodes were attached to the EEG amplifier. The participant was then instructed to sit still while EEG was recorded. The subjects were told to keep their eyes open during the measurement. If necessary, additional gel was applied to poorly behaving electrodes, and the electrode offset was checked to ensure proper connection. All offsets were ideally between +20 and −20 mV.

The EEG recordings were digitized at a sampling rate of 8192 Hz and stored on a hard disk using the BioSemi ActiView software, version 7.07 Before digitization, the activeTwo system applies a 5th-order cascaded integrator-comb (CIC) digital filter with a cut-off frequency of 1600 Hz

2.3.2. Digitizer

We acquired a 3D scan of the configuration of the EEG caps for all participants, using a Polaris Krios scanner (NDI; Waterloo ON, Canada), which scans all the electrodes, using a probe to mark three reference points: at the nasion and the height of the tragus at both sides. The Polaris Krios scanner is based on optical measurement technology and uses light reflected by markers to determine the position coordinates.

2.4. Stimulus Preparation

All stimuli were stored at a sampling rate of 48 kHz. For each stimulus file, a trigger file was generated. These triggers were sent from the stimulation equipment (Fireface UC, RME-audio; Haimhausen, Germany) to the BioSemi system. Triggers were generated every second in the form of a block wave. At every second and the beginning and end of the recording, a block pulse with a width of 1 ms was inserted. Based on the stimulus, speech-shaped noise was created at the same root-mean-square value (RMS) as the stimulus. The noise was created by taking white noise and changing the spectrum of the white noise to the spectrum of the speech and then matching the RMS value of the original stimulus file.

Afterward, using one noise file for each RMS value, the stimuli were calibrated with a type 2260 sound-level pressure meter, a type 4189 0.5 in. microphone, and a 2 cm³ coupler (Bruel & Kjaer, Copenhagen, Denmark).

The auditory stimuli were presented using a laptop connected to a Fireface UC (RME-audio; Haimhausen, Germany) using the APEX software platform, version 4.0.1 [41] and electromagnetically shielded insert phones (ER-3A, Etymotic; Fort Worth, TX, USA), binaurally at 62 dBA for each ear.

2.5. Preprocessed Data

Besides the raw EEG recordings, we also provide EEG and speech stimuli with commonly used preprocessing steps applied. All steps were conducted in Python 3.6, and the code for preprocessing is available on our GitHub repository (https://github.com/exporl/auditory-eeg-dataset (accessed on 1 June 2024)).

2.5.1. EEG

First, EEG data were high-pass filtered, using a 1st-order Butterworth filter with a cut-off frequency of 0.5 Hz. Zero-phase filtering was conducted by filtering the data forward and backward. Subsequently, the EEG was downsampled from 8192 Hz to 1024 Hz and eye-blink artifact removal was applied to the EEG, using a multichannel Wiener filter [42]. Afterward, the EEG was re-referenced to a common average, and finally, the EEG was downsampled to 64 Hz. The bandwidth of the resulting signal was therefore 0.5–32 Hz, as no low-pass filtering was performed (excluding the anti-aliasing filtering as performed by scipy.signal.resample_poly).

2.5.2. Speech Stimuli

The initial sampling frequency of the stimuli was 48 kHz. We provide a script to calculate the envelope using a gammatone filterbank [43] with 28 subbands. Each subband envelope was calculated by taking the absolute value of each sample, raised to the power of 0.6. A single envelope was obtained by averaging all these subbands [44]. Then, the envelope was downsampled to 64 Hz.

2.6. Validation

In order to demonstrate the validity of the data, we conducted several experiments on the preprocessed version of the proposed dataset, as explained in more detail in this section. All our results are reproducible using the code on our GitHub: https://github.com/exporl/auditory-eeg-dataset (accessed on 1 June 2024).

For all our experiments, we split each trial into a training, validation, and test set, containing, respectively, 80%, 10%, and 10% of each trial for each participant. The training, validation, and test sets did not overlap, so the test set remained unseen for all the models.

Before usage, we normalized each trial by computing the mean and standard deviation for each of the 64 EEG channels and the envelope stimulus on the training set. We then normalized the training, validation, and test sets by subtracting from each trial the mean and dividing by the standard deviation computed on the training set.

2.6.1. Linear Forward/Backward Modeling

To show the validity of the data, we trained participant-specific linear forward and backward models [1,2] (i.e., models that predict EEG from the stimulus envelope and the stimulus envelope from the EEG, respectively). The backward model was used to detect neural tracking in each recording, i.e., that the speech envelope could effectively be decoded for each participant/story compared to a null distribution of random predictions. The forward model was used to visualize the EEG channels for which the stimulus-related activity could be best predicted.

Model Training

The models were trained based on the recommendations of Crosse et al. (2021) [31]. The backward model weights were obtained similarly by Equation (1):

$$w_b={(R^{T}R+\lambda I)}^{-1}{R}^{T}s$$

(1)

where R is a matrix consisting of time-lagged versions of the EEG, s is the stimulus envelope, and $\lambda$ is the ridge regression parameter. In a similar fashion, Equation (2) was used to obtain the forward model weights:

$$w_f={(S^{T}S+\lambda I)}^{-1}{S}^{T}r$$

(2)

where S is a matrix consisting of time-lagged versions of the stimulus envelope, r is a matrix containing the EEG response, and $\lambda$ is the ridge regression parameter.

Both models utilized time lags from −100 ms to 400 ms. Following the recommendations of Crosse et al. (2021) [31], a leave-one-out cross-validation was performed on the recordings in the training set to determine the optimal ridge regression parameter ($\lambda$) from a list of values (${10}^x$ for $x=[-6,-4,-2,0,2,4,6]$). Correlation scores were averaged across folds and channels, after which the $\lambda$ was chosen, corresponding to the highest correlation value.

To evaluate the performance of both models, the Pearson correlation between the predicted and true data was calculated on the test set. In order to detect neural tracking, we followed the procedure of Crosse et al. (2021) [31]. For each recording in the test set, the predictions were (circularly) shifted in time by a random amount $N=100$ times. By correlating these shifted predictions to the actual signal, a null distribution was constructed for each participant. The 95th percentile of this null distribution was compared to the mean of the obtained scores on the test sets.

The analysis of EEG neural responses is typically performed in specific filter bands. For auditory EEG, the research typically focuses on the delta band (0.5–4 Hz) and the Theta band (4–8 Hz) [2,45,46]. We investigated the effect of filtering the EEG and envelope in different bands: delta (0.5–4 Hz), Theta (4–8 Hz), Alpha (8–14 Hz), Beta (14–30 Hz), and Broadband (0.5–32 Hz). A 1st-order Butterworth filter was chosen for each of the proposed filtering bands.

The model training and evaluation were performed in Python using Numpy [47] and Scipy.

Analysis

Using the linear backward model, we were able to detect neural tracking for all participants. In 11 of the 666 recordings, we were not able to detect neural tracking in any frequency band with the linear decoder. These recordings are listed in

Table 3. Recordings where no significant tracking was found with the linear backward model.

Subject	Stimulus
sub-002	audiobook_1_artefact
sub-011	audiobook_6_1
sub-051	audiobook_12_1
sub-051	audiobook_12_2
sub-051	podcast_23
sub-054	audiobook_12_2
sub-056	podcast_22
sub-060	podcast_24
sub-064	audiobook_14_2
sub-064	podcast_30
sub-076	audiobook_14_1

. The results per frequency band are shown in Figure 5. As previously shown by Vanthornhout et al. [2], the optimal performance was reached when filtering in the delta band (0.5–4 Hz). While correlations are hard to compare between studies because they are heavily influenced by the measurement paradigm, subject selection, preprocessing, and modeling choices, the correlations we found for the delta band were roughly in line with previous studies (median correlation between 0.1 and 0.2 [1,2]).

We compared the linear backward model performance across all stimuli and stimulus types (audiobooks vs. podcast, excluding the audiobook_1 shifted and artifact versions) in the delta band. The results are visualized in Figure 6 and Figure 7, respectively. Note that there was a large variability in decoding scores within and between stimuli. Additionally, a significant difference was found between the audiobook and podcast stimuli (0.184 vs. 0.133 median Pearson correlation, Mann-Whitney U test: $p<{10}^{-9}$). A similar observation was made for all the proposed architectures of the 2023 Auditory EEG decoding challenge [14], where the authors compared the performance of the models on stimulus type (audiobook vs. podcast) and observed that models performed significantly better on audiobooks than podcasts. A possible explanation for this might be that the podcasts are recorded using compression techniques, which generate an evenly loud signal. It might also be that the models fail to generalize to different compression techniques rather than to different unseen subjects.

Additionally, the backward model performance for the podcast with and without video is compared for the delta band in Figure 8. In that case, a decoder was trained separately for each recording on the training set with cross-validation and evaluated on the test set. No significant difference was found between the podcast presented with or without video. This is in contrast to previous studies that saw better neural tracking when the face of the speaker was presented [48,49]. A possible explanation for not finding an effect is that the face of the speaker was not continually present during the whole recording, removing the added benefit for at least some portion of the recording. Alternatively, the sample size contained within SparrKULee might be too small to find an effect across participants.

For the forward model, we show topomaps averaged across participants for each frequency band and stimulus type in Figure 9. As with the backward model, we observed the highest correlations between predicted and actual EEG signals in the delta band. The highest correlations were obtained for the channels in the temporal and occipital regions.

2.6.2. Non-Linear Models: Match-Mismatch Paradigm

For the non-linear models, we used the match-mismatch paradigm [7,12]. In this paradigm, the models were given three inputs: a segment of the EEG recording, the time-matched stimulus envelope segment, and a mismatched (imposter) stimulus envelope segment. As specified by Monesi et al. [10], the imposter was taken 1 s after the matched stimulus envelope segment. If extracting an imposter (at the end of each set) was impossible, the segment was discarded from the dataset. We extracted overlapping windows with 80% overlap. We included an analysis using a dilated convolutional model [12] to show typical match-mismatch performance across different input segment lengths.

Model Training

The dilated convolutional network consisted of four steps. First, the EEG channels were combined, from 64 to 8, using a 1D convolutional layer with a kernel size of 1 and a filter size of 8. Second, there were N dilated convolutional layers with a kernel size of K and 16 filters. These N convolutional layers were applied to both EEG and envelope stimulus segments. After each convolutional layer, a rectified linear unit (ReLU) was applied. Both stimulus envelope segments shared the weights for the convolutional layers. After these non-linear transformations, the EEG was compared to both stimulus envelopes, using cosine similarity. Finally, the similarity scores were fed to a single neuron, with a sigmoid non-linearity, to create a prediction of the matching stimulus segment.

The model was implemented in Tensorflow and used the Adam optimizer, with a learning rate of $0.001$ and binary-cross entropy as the loss function. Models were trained for a maximum of 50 epochs, using early stopping based on the validation loss, with a patience factor of 5. We trained the models with an input segment length of 5 s and in a participant-independent way, i.e., all participant data were given simultaneously to the model. We report results for input testing lengths of 1, 2, 3, 5, and 10 s. Since the trained dilation model did not have fixed input lengths, we used the same model with different input lengths.

Analysis

The results of this analysis can be seen in Figure 10. The accuracy of the model increased with longer window lengths. We see the same trend as in [12]. In order to test the generalizability of the model, we also tested the model with an arbitrarily chosen mismatch segment, as opposed to the fixed 1 s. There was no significant difference between these two testing conditions, which is in line with the experiment conducted in [50].

3. Data Description

All data were organized according to EEG-BIDS [51], an extension to the Brain Imaging Directory Structure (BIDS) [52] for EEG data. EEG-BIDS allows the storage of EEG data with relevant extra information files, e.g., about the experiment, the stimuli, and the triggers, enabling quick usage of the data and linking the auditory stimuli to the raw EEG files. A schematic overview of our repository is shown in Figure 11. The dataset consists of 3 parts: (1) raw data, in a folder per participant, (2) the auditory stimuli, in zipped Numpy (.npz) [47] format, and (3) the preprocessed data records, as described above, in the derivatives folder.

3.1. Raw Data

The raw data were structured as one folder per participant. For each participant (1 to 85), a folder sub-xxx is available in the root folder. In this folder, there is a folder indicating the session, which can be either ses-shortstories01 or ses-varyingstoriesxx ($xx=01\cdots 09$).

Each session folder contains a subfolder beh, containing the results of the behavioral matrix SRT estimation. These files were named according to the participant, the session, the task, which is always listeningActive, and the behavioral experiment run, which goes from one to three.

The data of the EEG experiment were stored in a subfolder in the session folder, named eeg. The EEG experiment data were named according to the participant, the session, the task, and the run. When the participant listened to a stimulus, the task was listeningActive. When the participant listened to silence, which happened at the start and end of the experiment, the task was restingState. The run suffix chronologically numbers the different trials starting at 01. Each trial has four corresponding files, differing only in their ending, after the run suffix: (1) raw gzipped file of EEG data in BioSemi Data Format (BDF), sampled at 8192 Hz, ending in eeg.bdf.gz; (2) a descriptive apr file eeg.apr, containing extra information about the experiment, such as the answers to the questions that were asked; (3) a stimulation file to link EEG to the corresponding stimulus stimulation.tsv; and (4) events.tsv, which describes which stimuli were presented to the participants at which time.

3.2. Stimuli

All the stimuli were saved in the folder stimuli/eeg. For each stimulus, we provide four corresponding files, stored in the npz format with additional gzipping to reduce storage, which is easily readable in Python: (1) the stimulus, stored at 48 kHz stimulusName.npz.gz, (2) the associated noise file noise_stimulusName.npz.gz, (3) the associated trigger file t_stimulueName.npz.gz, and (4) the experiment description file stimulusName.apx.

The stimuli were named according to their type: either audiobook_xx or podcast_xx, where xx indexes unique stimuli. Whenever an audiobook was split into multiple consecutive parts, an extra suffix denotes which part of the audiobook is referred to.

3.3. Preprocessed Data

For all data, we also provide a preprocessed, downsampled, and time-aligned version of the data. These data can be found in the derivatives/preprocessed_eeg folder. Similar to the raw data, the preprocessed data were structured in a folder per participant, per session, which could be either ses-shortstoriesxx or ses-varyingstoriesxx. The preprocessed files derived their name from the raw EEG file used to create the preprocessed version. To avoid confusion, a suffix desc-preproc was added, such that no two files have the same name. After the desc-preproc suffix, the name of the stimulus the participant listened to was added to facilitate linking the EEG brain response to the auditory stimulus for downstream tasks.