Paper deep dive
Cephalonauts One: A deep fMRI dataset for decoding naturalistic speech in the human brain
Antoine Collas, Louis Jalouzot, GÊraud Ilinca, Corentin Caris, Romain Valabrègue, Ahmed Hassayoune, David Goncalves, Madeleine Hueber, ThaddÊe Delebarre, Julien Savatovsky, Clara Fonteneau, Charles Maussion, Bertrand Thirion, Alexis Thual
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Cephalonauts One is a whole-brain 3 Tesla (3T) functional magnetic resonance imaging (fMRI) dataset recorded while subjects listened to audio podcasts. Three healthy subjects underwent multiple scanning sessions, each consisting of five 15-minute runs, while listening to podcasts in their native language. With 30 hours of fMRI data per subject, the current release is the deepest available fMRI dataset using naturalistic speech stimuli. The dataset pairs brain activity with the corresponding podcast audio, transcript annotations, and derived stimulus embeddings. Furthermore, we introduce a brain decoding benchmark formulated as audio segment retrieval: given fMRI activity from a held-out session, the decoder must identify the corresponding time-aligned podcast audio segment among candidate segments. We provide standardized splits, evaluation metrics, and baseline decoders for this task. Finally, a scaling analysis shows that decoding performance improves continuously with the amount of training data per subject.
Tags
Links
- Source: https://arxiv.org/abs/2610.03558v1
- Canonical: https://arxiv.org/abs/2610.03558v1
Trouble viewing inline? Open PDF directly â
Full Text
88,137 characters extracted from source content.
Expand or collapse full text
Cephalonauts One: A deep fMRI dataset for decoding naturalistic speech in the human brain Antoine Collas Affiliation: Karavela Email: acollas@karavela.ai Louis Jalouzot Affiliation: Karavela; UNICOG, CNRS, INSERM, CEA, Paris-Saclay University; LSCP, EHESS, ENS, CNRS, PSL University GĂŠraud Ilinca Affiliation: Karavela Corentin Caris Affiliation: Karavela Romain Valabrègue Affiliation: Centre de NeuroImagerie de Recherche (CENIR), Sorbonne UniversitĂŠ, ICM, Paris, France Ahmed Hassayoune Affiliation: Karavela David Goncalves Affiliation: Karavela Madeleine Hueber Affiliation: Karavela ThaddĂŠe Delebarre Affiliation: Karavela Julien Savatovsky Affiliation: Department of Radiology, HĂ´pital Fondation Adolphe de Rothschild, Paris, France Clara Fonteneau Affiliation: Karavela Charles Maussion Affiliation: Karavela Bertrand Thirion Affiliation: Karavela; Inria, CEA, Paris-Saclay University, Palaiseau, France Email: bthirion@karavela.ai Alexis Thual Affiliation: Karavela Email: athual@karavela.ai Abstract Cephalonauts One is a whole-brain 3 Tesla (3T) functional magnetic resonance imaging (fMRI) dataset recorded while subjects listened to audio podcasts. Three healthy subjects underwent multiple scanning sessions, each consisting of five 15-minute runs, while listening to podcasts in their native language. With 30 hours of fMRI data per subject, the current release is the deepest available fMRI dataset using naturalistic speech stimuli. The dataset pairs brain activity with the corresponding podcast audio, transcript annotations, and derived stimulus embeddings. Furthermore, we introduce a brain decoding benchmark formulated as audio segment retrieval: given fMRI activity from a held-out session, the decoder must identify the corresponding time-aligned podcast audio segment among candidate segments. We provide standardized splits, evaluation metrics, and baseline decoders for this task. Finally, a scaling analysis shows that decoding performance improves continuously with the amount of training data per subject. Data & code: huggingface.co/datasets/karavela/cephalonauts_one Demo: karavela.ai/en/research/cephalonauts-one Figure 1: Overview of the Cephalonauts One dataset. A. Three subjects listened to podcasts during fMRI acquisition. B. Each session included five 15-minute runs, three comprehension questions after each run, and a final anatomical scan. Runs 1â2, 3â4, and 5 used three distinct podcasts and podcasts were not repeated across sessions. 1 Introduction Brain-behavior links The human brain can process highly complex perceptual inputs but also produce rich, intentional, and reflective thoughts. Extensive research has been devoted to probing, measuring, and modelling this complex biological system, in the hope of gaining clinical insights into brain disorders, inspiring the design of more intelligent artificial systems, and better grasping the nature of human cognition (Mesulam, 1998). One way of gaining information about the brain is to link its properties â e.g., anatomical, functional, or biological â to properties of the external world â e.g., a list of tasks an individual is able or unable to perform, or the category of stimuli they are exposed to. Famous historical approaches demonstrated that lesions in specific frontal and temporal regions could impair speech and comprehension, yielding early insights into the functional organization of the language system (Broca, 1861; Wernicke, 1874). Brain encoding and decoding More recently, functional magnetic resonance imaging (fMRI) (Huettel et al., 2004) enabled non-invasive measurement of brain activity in healthy subjects. Its higher spatial and temporal resolution compared with positron emission tomography (PET) made it possible to characterize brain responses in greater detail. This led to early statistical analyses showing that some brain regions respond selectively to specific stimulus categories (Kanwisher et al., 1997). To identify regions where these effects cannot necessarily be observed so clearly, a large body of work has predicted recorded brain activity from known stimulus features, a process known as encoding. Encoding models characterize how stimulus features are represented across the cortex, from retinotopic organization in early visual areas (Dumoulin and Wandell, 2008) to semantic maps estimated from fMRI recorded while participants listened to long natural narratives (Huth et al., 2016). In parallel, many studies have attempted to infer properties of the stimulus from recorded brain activity, a process known as decoding. In language neuroscience, recent decoding approaches have moved beyond coarse stimulus labels by mapping non-invasive brain recordings to rich speech representations. Jalouzot et al. (2025) showed that fMRI decoding of natural speech improves with additional subject-specific data, a finding consistent with related speech-decoding results in magneto- and electroencephalography (M/EEG) (DĂŠfossez et al., 2023). Stimulus embeddings In fact, brain encoding and decoding can be viewed as two sides of the same coin, as both aim to link properties of brain activity to properties of the stimulus. While early work focused on relatively low-dimensional variables such as stimulus category or the presence of cognitive impairments, the advent of deep learning has made it possible to access high-level high-dimensional representations of the stimulusâso-called latent representations or embeddingsâthat can be used in both encoding (Eickenberg et al., 2017) and decoding analyses (DĂŠfossez et al., 2023). In both settings, these representations support increasingly expressive models with greater predictive power (dâAscoli et al., 2026), opening new perspectives for understanding brain function. Natural speech Language is a particularly compelling stimulus, spanning multiple levels from acoustics to semantics and narrative context. Spoken narratives are especially useful for studying these representations in the brain: they engage distributed cortical systems (Huth et al., 2016) over multiple timescales (Lerner et al., 2011), while remaining easy to align with transcripts and computational representations from modern large language models (LLMs) (Zhang et al., 2025; Khare et al., 2025). Prior work used brain responses to compare natural language processing embeddings, showing that they capture different aspects of language (Toneva and Wehbe, 2019). Other studies have shown that modern models partly match human brain responses (Caucheteux and King, 2022), or used controlled models to separate the effects of meaning, syntax, and context in the cortex (Pasquiou et al., 2023). Together, these results show that fMRI during speech listening can help study both biological and computational language representations. Dataset depth However, fMRI with speech stimuli remains constrained by limited per-participant data. This is limiting for subject-specific decoders, for which performance improves with additional within-subject data in both natural speech and visual fMRI decoding (Tang et al., 2023; Scotti et al., 2023). Indeed, large-scale analyses further indicate that decoding scales primarily with data per subject (Banville et al., 2025). Existing speech fMRI resources address different regimes: Narratives provides up to roughly 2 hours of unique stimuli per participant (Nastase et al., 2021), whereas LeBel et al. (2023) provide up to 16 hours for 3 participants. Hence, deep per-subject language fMRI datasets for natural speech encoding and decoding remain scarce. 2 Related work Table 1: Brain datasets using speech stimuli. Maximum hours per subject reports the maximum subject-specific stimulus duration. For LeBel et al. (2023) we report the 2 available data regimes. Benchmark-oriented denotes releases with benchmark tasks, metrics, baselines, and code. Dataset Modality Nb. of subjects Max hours per subject Benchmark- oriented References Cephalonauts One fMRI 3 30 Yes This work LibriBrain MEG 1 52.3 Yes Ăzdogan et al.,2025 LeBel et al. fMRI 8 / 3 6 / 16 No LeBel et al.,2023 Le Petit Prince fMRI 112 1.7 No Li et al.,2022 Armeni et al. MEG 3 10 No Armeni et al.,2022 Narratives fMRI 345 2 No Nastase et al.,2021 Early scaling of naturalistic fMRI datasets Public neural datasets have been central to progress in brain encoding and decoding. Early scaling was driven by fMRI datasets based on visual stimuli. Indeed, early decoding datasets contained roughly 1.5â3 stimulus hours per subject (Wen et al., 2017; Chang et al., 2019). Then, the Natural Scenes Dataset increased visual stimuli to 40 hours per subject (Allen et al., 2022). More recently, the Courtois NeuroMod datasets pushed further with roughly 80 hours of movie watching per subject (Boyle et al., 2025). This scale has supported advances in visual encoding and decoding, including natural-image reconstruction from fMRI with deep learning models (Chen et al., 2022; Ozcelik and VanRullen, 2023; Scotti et al., 2023). Speech stimuli in fMRI datasets: breadth and depth However, speech fMRI datasets have often scaled across rather than within participants. Narratives aggregates story-listening fMRI from 345 participants, with roughly 4.6 hours of unique audio stimuli across the full dataset (Nastase et al., 2021), while Le Petit Prince provides multilingual audiobook fMRI from 112 subjects, with approximately 1.7 hours per subject (Li et al., 2022). Shared stimuli support group analyses and cross-subject comparisons, but limited per-subject depth constrains subject-specific decoders. Recently, LeBel et al. (2023) went deeper with natural story-listening fMRI from 8 participants, including an extended subset of up to 16 hours of podcast listening for 3 participants. Using these data, Jalouzot et al. (2025) show that natural speech decoding improves with more training data per participant. Benchmark-ready datasets In an effort to make brain encoding and decoding more reproducible and accessible, recent datasets provide standardized tasks, code, and baseline models. The Algonauts Project 2025 Challenge (Gifford et al., 2024) turns CNeuroMod movie fMRI data into an encoding benchmark with private in-distribution and out-of-distribution test sets, public leaderboards, and a baseline model. Similar releases exist for other recording modalities: LibriBrain (Ăzdogan et al., 2025) provides 52 hours from one subject, standardized splits, public code, and baseline speech-decoding tasks. Together, such datasets support model comparison and progress tracking. Remaining gap fMRI provides whole-brain spatial coverage and a better decoding-performance-to-stimulus-duration ratio than M/EEG (Banville et al., 2025). However, as summarized in Table 1, brain decoding still lacks fMRI datasets that are both deep at the subject level and released as ready-to-use benchmarks. Such datasets should provide aligned stimulus representations, held-out evaluation sessions, standardized metrics, and reproducible baseline decoders. Contributions We introduce Cephalonauts One, a deep per-subject dataset for benchmarking decoding models of naturalistic speech from fMRI. The main contributions are: ⢠A deep fMRI dataset whose current release comprises 30 hours per subject of whole-brain 3T fMRI from 3 subjects listening to naturalistic speech; ⢠Paired brain and stimulus data linking preprocessed fMRI activity to podcast audio, transcripts, and derived stimulus embeddings; ⢠A brain decoding benchmark for retrieving podcast audio segments from held-out fMRI activity, with standardized splits, metrics, and reproducible baselines; ⢠A decoding scaling analysis showing that performance improves approximately linearly on a logâlog scale with per-subject data, with no saturation over the observed range. 3 The Cephalonauts One dataset We release preprocessed fMRI recordings paired with stimulus representations to support the development and benchmarking of brain decoding and encoding models. The dataset and baseline code are hosted at huggingface.co/datasets/karavela/cephalonauts_one, where the repository README provides additional loading instructions, decoding baselines, and usage details. After authentication and access approval, the data can be loaded with the Hugging Face datasets package: from datasets import load_dataset dataset_name = "karavela/cephalonauts_one" subset_name = "sub-0_fsaverage5" dataset = load_dataset(dataset_name, name=subset_name, split="train", streaming=True) row = next(iter(dataset)) print(list(row)) # ["timestamps", "text", "X", "Y_text", "Y_audio", "sub", "ses", "run", # "tr", "fmri_offset", "audio", "audio_offset", "audio_duration", # "run_fmri_duration", "fmri", "split"] print("fMRI dim:", len(row["X"]), "audio embedding dim:", len(row["Y_audio"])) # fMRI dim: 20484, audio embedding dim: 3584 The fMRI data are provided in multiple spatial representations, summarized in Table 3, including both volumetric and cortical surface representations, each released on Hugging Face as a separate subjectâspace subset. We also provide a separate timeseries subset with run-level preprocessed time series, masks, transcripts, and audio files, allowing users to rebuild the aligned data or define new analyses. The stimulus data are likewise released in multiple embedding formats derived from both text and audio, as detailed in Table 4. 3.1 Dataset overview Table 2: Dataset size by subject. Subject # TRs # sessions # runs Duration (h) sub-0 207,283 24 120 29.94 sub-1 209,366 24 121 30.24 sub-2 201,647 24 117 29.13 Total 618,296 72 358 89.31 Acquisition protocol As illustrated in Figure 1, Cephalonauts One consists of fMRI recordings from three subjects while they listened to podcasts in their native language, French. Each session comprised approximately 75 minutes of functional scanning, together with one anatomical T1-weighted scan. The dataset contains 30 hours of functional recordings per subject (see Table 2). Within each subject, stimuli were never repeated, although some stimuli were shared across subjects. Each functional session consisted of five 15-minute runs organized around three distinct podcasts: runs 1â2 and 3â4 presented two podcasts split across consecutive runs, while run 5 presented a third podcast. For each run, the timing of audio onset relative to MRI acquisition onset was recorded to synchronize the stimulus with the fMRI time series. Participants were instructed to listen attentively to each podcast run. After each run, they answered three multiple-choice comprehension questions about the story they had just heard. Each question had three candidate answers, from which the subject selected one. These questions were generated automatically using Gemini 3 Pro (Gemini Team and others, 2024). Three healthy adult volunteers, aged 31â51 years, were recruited for this study. All participants were native French speakers, had no history of neurological or psychiatric disorders, and met standard MRI safety criteria. The study was approved by an ethics committee (CPP Ile de France 2 2025-A02302-47 on February 16, 2026), with written informed consent given by all participants and monetary compensation provided (80 ⏠per session). Stimuli Subjects listened to Les Pieds sur Terre11 1 https://w.franceculture.fr/emissions/les-pieds-sur-terre, a popular French podcast produced by the public radio station France Culture that features real-life stories told by ordinary people. Episodes are approximately 30 minutes long and center on first-person true stories about themes such as everyday life, relationships, work, personal turning points, or social change. Each episode typically features a single main speaker, who narrates a personal experience in rich and diverse language. Although the same journalist generally introduces the episodes, the story itself is usually delivered as a monologue by the featured speaker. The narration is often interspersed with short musical interludes. 3.2 Brain activity recordings Table 3: Released brain data. We release three versions of the preprocessed fMRI data generated with DeepPrep: native T1w space, MNI space, and the fsaverage5 cortical surface. V denotes the number of voxels or surface vertices. Space Vol. Surf. Subject anat. Template anat. V Description T1w Yes No Yes No âź 50,000 whole-brain in each subjectâs native anatomical space MNI Yes No No Yes 69,765 whole-brain normalized to 3 m MNI152NLin6Asym standard space fsaverage5 No Yes No Yes 20,484 both hemispheres on the standard fsaverage5 surface MRI acquisition Data were acquired at the Paris Brain Institute on a 3T Siemens MAGNETOM Cima.X whole-body MRI scanner equipped with a 64-channel head/neck coil. Functional images were collected with a highly accelerated 2D gradient-echo echo-planar imaging (EPI) sequence using simultaneous multi-slice acquisition with multiband factor 6, yielding a repetition time (TR) of 520 ms. The protocol was designed to provide high temporal resolution while maintaining whole-brain coverage. Functional images were acquired at 3 m isotropic resolution. Each run contained âź 1700 functional brain volumes. High-resolution anatomical images were acquired with a 3D T1-weighted spoiled gradient-echo Dixon sequence at 0.8 m isotropic resolution. The same acquisition protocol was used across all sessions and subjects. Each subject underwent 30 hours of fMRI scanning, yielding âź 207,000 brain volumes. Additional sequence details are provided in Appendix S1.3, including the core acquisition parameters summarized in Table S1. Preprocessing Functional images were preprocessed with DeepPrep (Ren et al., 2025), a fully automated fMRI preprocessing pipeline that includes standard spatial preprocessing steps such as motion correction, spatial normalization, and anatomical-to-functional alignment. DeepPrep uses FreeSurfer (Fischl, 2012) for anatomical processing, including cortical surface reconstruction and segmentation. As summarized in Table 3, we release the resulting data in three complementary spaces: subject-native T1w space, MNI152NLin6Asym space, and the fsaverage5 cortical surface. These representations support different use cases: subject-native T1w space preserves individual anatomy for within-subject analyses, MNI space facilitates cross-subject comparisons and compatibility with standard atlases, and fsaverage5 focuses the input on cortical gray matter while reducing dimensionality. Broad cortical representations are particularly relevant for natural speech fMRI, since semantic responses during story listening are distributed across large regions of cortex (Huth et al., 2016). Surface projections were computed from each subjectâs anatomical scan to ensure anatomically consistent mapping from volumetric data to cortex. For the volumetric T1w and MNI spaces, we apply the corresponding brain mask before export, so each run is stored as a lighter and more decoding-ready two-dimensional time-by-voxel array rather than as a 4D image; the masks are also released to map these arrays back to brain space. For decoding analyses, each voxel or vertex time series is additionally detrended, z-scored, and temporally smoothed with a moving-average filter of width 4 TRs. 3.3 Stimulus Stimulus preprocessing For each run, we release the original stimulus in two modalities: the raw audio signal and its transcript. Word-level transcripts were generated automatically with AssemblyAI Universal 3 Pro. We converted the word-level transcript into a sequence of regularly sampled stimulus onsets spaced every 4 TRs, corresponding to 2.08 seconds. At each onset, we extracted the immediately preceding context by collecting the last 20 completed words whose offsets occurred before that time, and used their concatenation as the text stimulus. For the audio modality, we used the same retained onsets and represented each event by the 10-second audio segment ending at that onset. Early onsets with insufficient preceding context were discarded. As summarized in Table 4, we release both these segment-level data and the corresponding embeddings. Stimulus embeddings We embedded text segments with Qwen3-Embedding-8B and audio segments with LCO-Embedding-Omni-7B, yielding 4096-dimensional and 3584-dimensional representations, respectively (Zhang et al., 2025; Xiao et al., 2025). This choice is motivated by the fact that they are leading models on their respective benchmarks: MTEB multilingual for text (Muennighoff et al., 2023) and MAEB for audio (Assadi et al., 2026). These benchmarks evaluate general-purpose embedding spaces across diverse retrieval and similarity tasks, matching our use of text and audio embeddings as targets for held-out segment retrieval. Alignment to fMRI Each retained stimulus event was paired with the corresponding fMRI measurement by selecting the brain frame acquired after the event, accounting for the hemodynamic lag, set to 2.5 seconds. This fixed lag approximates the delay between stimulus processing and the measured BOLD response. This procedure yields aligned pairs of brain volumes and stimulus embeddings for any machine learning tasks, as illustrated in Figure 2.A. Table 4: Released stimulus representations. For each stimulus modality, we release both segment-level data and embeddings from foundation models. D denotes the embedding dimensionality. Modality Representation D Description Text segments â overlapping 20-word transcript contexts sampled at retained onsets embeddings 4,096 embeddings of transcript segments computed with Qwen3-Embedding-8B Audio segments â 10-second audio segments ending at retained onsets embeddings 3,584 embeddings of audio segments computed with LCO-Embedding-Omni-7B 4 Decoding experiments Cephalonauts One provides a benchmark for decoding naturalistic speech from fMRI. The task is formulated as audio segment retrieval: given brain activity from a held-out session, the model must identify the corresponding podcast audio segment among all other possible segments from that session. This task fixes the evaluation target while leaving the decoding approach open: models may use any stimuli representation, decoder architecture, or training objective, provided they are evaluated on the same held-out retrieval problem. The proposed baseline predicts embeddings of the stimulusâ raw audio and transcript from fMRI activity and ranks candidate segments by cosine similarity in that embedding space. Figure 2 presents the overall framework. 4.1 Experimental setup Figure 2: Dataset construction and decoding framework. A. For each frame t, stimulus t z_t is aligned with fMRI recordings sampled every 520 ms. Target stimulus embeddings t y_t summarize the audio waveform or transcript over the stimulus context leading up to t. The corresponding brain input uses preprocessed fMRI volumes shifted by a fixed lag. In this illustration, the stimulus context spans 1.56 seconds (3 TRs) and the lag is set to 1.04 seconds (2 TRs). In the experiments, we set the lag to 2.5 seconds and the stimulus context to 10 seconds. B. A pretrained encoder maps stimulus Z to stimulus embeddings Y. The decoder fθf_θ is trained to predict Y from pre-processed brain recordings X. Decoding task For each subject, we evaluate a subject-specific decoder on five folds, using session indices 19â23 once each for test. Each fold holds out the next session in this set for validation (wrapping from 23 to 19); the remaining 22 sessions form the training pool. This split ensures that validation and test samples come from recording sessions and podcast segments that were not used for training to more accurately measure generalization. At test time, the decoder receives an fMRI frame and has to retrieve the corresponding podcast audio segment among approximately N=2,000N=2,000 candidates from the held-out session. The candidates therefore come from unseen podcast segments in the same session, making retrieval difficult because many segments share the same speaker, topic, and context. Evaluation metrics For each test sample i, the decoder predicts an embedding ^i y_i from fMRI activity. The N stimulus embeddings from the held-out session, jj=1N\ y_j\_j=1^N, are used as retrieval candidates. We compute cosine similarities siâjs_ij between the predicted embedding and each candidate, and define rir_i as the rank of the true aligned stimulus, i.e. how many other candidates are closer to the prediction than the true stimulus is: siâjâ^iâ¤âjâ^iâ2ââjâ2,riââj=1N[siâj>siâi]+1,s_ij\; \; y_i y_j\| y_i\|_2\;\| y_j\|_2, r_i\; \; _j=1^N1 [s_ij>s_i ]+1, (1) where â[â ]1[¡] is 1 when the condition holds, 0 otherwise. We report three metrics. Top-10@N is the fraction of test samples whose true stimulus is among the top 10 candidates: Top-10@Nâ1Nâi=1N[riâ¤10].Top-10@N\; \; 1N _i=1^N1 [r_i⤠10 ]. (2) Median Relative Rank measures normalized median rank of the true stimulus: Median Relative RankâmedianâĄ(riNâŁiâ1,âŚ,N).Median Relative Rank\; \;median ( \ r_iN iâ\1,âŚ,N\ \ ). (3) Finally, Cosine Similarity measures the similarity between the predicted embedding and the true stimulus embedding, averaged across test samples: Cosine Similarityâ1Nââi=1Nsiâi.Cosine Similarity\; \; 1N _i=1^Ns_i. (4) These metrics capture complementary aspects of decoding performance. Top-10 Retrieval Accuracy measures whether the decoder places the correct stimulus among a small set of plausible candidates. Median Relative Rank summarizes the ranking distribution and is not sensitive to a single arbitrary cutoff nor to the retrieval set size. Cosine Similarity measures the quality of the predicted embedding directly, independently of the candidate pool used for retrieval. Together, these metrics distinguish accurate embedding prediction from successful retrieval among held-out stimulus segments. Baseline decoder We provide a simple decoder baseline for this retrieval task. The decoder is a one-hidden-layer neural network trained to predict a stimulus embedding from brain activity. Each target embedding is defined as the concatenation of the text and audio embeddings described in Section 3.3. We train the decoder with AdamW (Kingma and Ba, 2017; Loshchilov and Hutter, 2017) using the CLIP contrastive loss (Radford et al., 2021) with temperature 0.02, following recent brain decoding setups (Scotti et al., 2023; Jalouzot et al., 2025). Training uses early stopping on the validation Median Relative Rank, with patience of five epochs, and the best validation checkpoint is used for test evaluation. At evaluation time, candidate segments are ranked by cosine similarity between their embeddings and the predicted embedding. All hyperparameters are reported in Appendix S2. 4.2 Results Figure 3: Speech decoding performance versus training-set size. Test performance is shown versus training hours for each subject, using Median Relative Rank, Top-10@2000, and Cosine Similarity. Lower is better for Median Relative Rank; higher is better for the other metrics. Error bars show one empirical standard deviation across five folds with distinct held-out test sessions. Performance improves with additional subject-specific data, with no clear saturation. Chance level is 0.5 for Median Relative Rank and âź 0.5%\% for Top-10@2000. Performance scales with training data Figure 3 shows decoding performance as a function of training-set size. Across subjects, performance improves as more subject-specific training sessions are added. Cosine Similarity increases, Median Relative Rank decreases, and Top-10 Retrieval Accuracy increases over the range of training sizes. For example, Cosine Similarity increases from approximately 0.075â0.090 at the smallest training sizes to approximately 0.144â0.155 at the largest training sizes. Over the same range, Median Relative Rank decreases from approximately 0.11â0.19 to 0.022â0.034, while Top-10 Retrieval Accuracy increases from approximately 3.8â6.4% to 22.3â27.3%. These gains are visible across all three metrics, indicating that the decoder consistently benefits from additional within-subject recordings. As in the empirical scaling laws reported for neural language models (Kaplan et al., 2020), the trajectories are approximately linear in the log-log space of Figure 3, with performance continuing to improve up to the largest training sets and no clear saturation over the range studied. This supports the value of deep per-subject datasets for naturalistic language decoding. 5 Discussion Cephalonauts One is designed as both a neuroscience resource and a machine-learning benchmark for naturalistic language decoding. Its main contribution is to increase within-subject depth for speech fMRI, while pairing brain activity with aligned text and audio representations. This design creates opportunities for studying language processing, evaluating multimodal decoding models, and developing methods that may later transfer to clinical settings. At the same time, the dataset has important limitations: it includes only three subjects, focuses on French podcast stimuli, and evaluates decoding under a controlled retrieval setting. We discuss these broader impacts and limitations below. 5.1 Broader impact Neuroscience Cephalonauts One can support a better understanding of how the human brain processes naturalistic speech. By pairing many hours of fMRI with aligned audio and transcript representations, the dataset makes it possible to study language comprehension over extended timescales, from acoustic input to semantic content. Its depth also supports subject-specific analyses. Multimodal machine learning Cephalonauts One also provides a useful setting for multimodal representation learning. Text and audio benefit from strong pretrained foundation models, whereas fMRI does not yet have comparably mature foundation models. This creates an asymmetric multimodal problem: one modality has rich pretrained representations, while the other remains noisy, high-dimensional, and data-limited. The benchmark therefore supports the training and evaluation of models that align brain activity with established text and audio embedding spaces. More broadly, it can help study how machine learning systems handle modalities with unequal levels of representation quality, data availability, and pretrained model support. Potential negative impacts Brain recordings are sensitive biomedical data, and decoding cognition could raise privacy concerns. To best protect participants privacy, access to Cephalonauts One is gated on Hugging Face and requires users to accept terms of use before requesting access. Participants gave informed consent for data sharing, and the terms of use prohibit re-identification attempts, contact with study volunteers, and provision of information linking the data to protected medical information. 5.2 Limitations Limited number of subjects Cephalonauts One prioritizes depth over breadth. It contains many hours of recordings per subject, but only three subjects. This design supports subject-specific decoding and scaling analyses, but limits conclusions about population-level variability and cross-subject generalization. Single-site acquisition All data were collected at the same imaging site, on the same scanner, with the same MRI sequence. This improves consistency across sessions and subjects, but limits conclusions about robustness to other sites, scanners, field strengths, and acquisition settings. MRI acquisition sequence The acquisition protocol was optimized for temporal sampling rather than spatial resolution. The 520 ms TR provides many brain volumes per hour of listening. This choice came with a 3 m isotropic functional resolution, which is lower than protocols designed for fine-grained spatial mapping. Thus, Cephalonauts One is well suited to studying temporally dense naturalistic speech decoding, but it is less suited to questions that require precise localization of small cortical structures. Restricted language and stimulus domain All subjects listened to French podcasts from a single program, Les Pieds sur Terre. This gives coherent and naturalistic stimuli, but restricts the dataset to French narrative speech. Models trained on Cephalonauts One may not generalize to other languages or broader auditory environments. Retrieval-based evaluation Our benchmark focuses on retrieval: the model must identify the correct audio segment from a finite pool of candidates in a held-out session. This setting is already challenging because candidates come from unseen podcast segments with similar topic, speaker, and narrative context. However, it does not cover the full range of possible decoding tasks. We hope that Cephalonauts One will also support future benchmarks at different levels of difficulty, from easier tasks such as speech-versus-non-speech or coarse stimulus classification to harder tasks such as open-ended generation of text or audio stimuli from brain activity. 6 Conclusion Cephalonauts One provides 30 hours of whole-brain 3T fMRI per subject from three subjects listening to French podcasts, paired with aligned audio and transcript representations. It includes preprocessed brain data in multiple spaces, stimulus embeddings from text and audio, standardized decoding tasks, baseline models, and scaling analyses over training duration. Decoding performance improves with additional within-subject data, with no clear saturation over the range studied, supporting deep recordings for naturalistic language decoding. Acknowledgments and Disclosure of Funding The authors thank Julie Dessoubrais, Brunhilde Ventre, Margaux Couppey, Chahrazad Chenaf, Marine Kavalko, Marjorie Notolan, Silvia Adins, StĂŠphanie Anastacio, Marie Ropiquet, Jasmina Wallace, Marc Lapert, BenoĂŽt BĂŠranger, ClĂŠment Barbe, MĂŠlanie Didier, Ăric Bardinet, and StĂŠphane LehĂŠricy for their help with data collection. This work was funded by Karavela and benefited from HPC resources provided by GENCIâIDRIS (Grant 2025-AD010316590R1). The authors declare no additional competing interests beyond their listed primary affiliations. References Abraham et al. (2014) A. Abraham, F. Pedregosa, M. Eickenberg, P. Gervais, A. Mueller, J. Kossaifi, A. Gramfort, B. Thirion, and G. Varoquaux Machine learning for neuroimaging with scikit-learn. Frontiers in Neuroinformatics 8. External Links: ISSN 1662-5196, Link Cited by: Appendix S3. Allen et al. (2022) E. J. Allen, G. St-Yves, Y. Wu, J. L. Breedlove, J. S. Prince, L. T. Dowdle, M. Nau, B. Caron, F. Pestilli, I. Charest, J. B. Hutchinson, T. Naselaris, and K. Kay A massive 7T fMRI dataset to bridge cognitive neuroscience and artificial intelligence. Nature Neuroscience 25 (1), p. 116â126 (en). External Links: ISSN 1546-1726, Link, Document Cited by: §2. Armeni et al. (2022) K. Armeni, U. GßçlĂź, M. van Gerven, and J. Schoffelen A 10-hour within-participant magnetoencephalography narrative dataset to test models of language comprehension. Scientific Data 9 (1), p. 278. Cited by: Table 1. Assadi et al. (2026) A. E. Assadi, I. Chung, C. Xiao, R. Solomatin, A. Jha, R. Chand, S. Singh, K. Wang, A. S. Khan, M. M. Nasser, et al. MAEB: massive audio embedding benchmark. arXiv preprint arXiv:2602.16008. Cited by: §3.3. Banville et al. (2025) H. Banville, Y. Benchetrit, S. dâAscoli, J. Rapin, and J. King Scaling laws for decoding images from brain activity. arXiv. Note: arXiv:2501.15322 [eess] External Links: Link, Document Cited by: §1, §2. Boyle et al. (2025) J. A. Boyle, B. Pinsard, L. (. Bellec, and The CNeuroMod team CNeuroMod Data Collection Complete: 200h of individual fMRI Across Diverse Naturalistic and Controlled Tasks to build NeuroAI models. In Conference on Cognitive Computational Neuroscience (CCN), (en). Note: Conference abstract External Links: Link Cited by: §2. Broca (1861) P. Broca Perte de la parole, ramouissement chronique et destruction partielle du lobe antĂŠrieur gauche du cerveau. Bulletins de la SociĂŠtĂŠ Anthropologique de Paris 2, p. 235â238. External Links: Link Cited by: §1. Caucheteux and King (2022) C. Caucheteux and J. King Brains and algorithms partially converge in natural language processing. Communications Biology 5 (1), p. 134 (en). Note: Number: 1 External Links: ISSN 2399-3642, Link, Document Cited by: §1. Chang et al. (2019) N. Chang, J. A. Pyles, A. Marcus, A. Gupta, M. J. Tarr, and E. M. Aminoff BOLD5000, a public fMRI dataset while viewing 5000 visual images. Scientific Data 6 (1), p. 49 (en). External Links: ISSN 2052-4463, Link, Document Cited by: §2. Chen et al. (2022) Z. Chen, J. Qing, T. Xiang, W. L. Yue, and J. H. Zhou Seeing Beyond the Brain: Conditional Diffusion Model with Sparse Masked Modeling for Vision Decoding. arXiv. Note: arXiv:2211.06956 [cs] External Links: Link, Document Cited by: §2. Ciric et al. (2022) R. Ciric, W. H. Thompson, R. Lorenz, M. Goncalves, E. E. MacNicol, C. J. Markiewicz, Y. O. Halchenko, S. S. Ghosh, K. J. Gorgolewski, R. A. Poldrack, et al. TemplateFlow: fair-sharing of multi-scale, multi-species brain models. Nature methods 19 (12), p. 1568â1571. Cited by: Appendix S3. da Costa-Luis et al. (2026) C. da Costa-Luis, S. K. Larroque, K. Altendorf, et al. tqdm: a fast, extensible progress bar for Python and CLI. Zenodo. Note: Software archive, all versions External Links: Document, Link Cited by: Appendix S3. DĂŠfossez et al. (2023) A. DĂŠfossez, C. Caucheteux, J. Rapin, O. Kabeli, and J. King Decoding speech perception from non-invasive brain recordings. Nature Machine Intelligence 5 (10), p. 1097â1107 (en). External Links: ISSN 2522-5839, Link, Document Cited by: §1, §1. Dumoulin and Wandell (2008) S. O. Dumoulin and B. A. Wandell Population receptive field estimates in human visual cortex. Neuroimage 39 (2), p. 647â660. Cited by: §1. dâAscoli et al. (2026) S. dâAscoli, J. Rapin, Y. Benchetrit, T. Brooks, K. Begany, J. Raugel, H. J. Banville, and J. King A foundation model of vision, audition, and language for in-silico neuroscience. External Links: Link Cited by: §1. Eickenberg et al. (2017) M. Eickenberg, A. Gramfort, G. Varoquaux, and B. Thirion Seeing it all: Convolutional network layers map the function of the human visual system. NeuroImage 152, p. 184â194. External Links: Link Cited by: §1. Fischl (2012) B. Fischl FreeSurfer. NeuroImage 62 (2), p. 774â781 (en). External Links: ISSN 1053-8119, Link, Document Cited by: Appendix S3, §3.2. Gemini Team et al. (2024) Gemini Team et al. Gemini: A Family of Highly Capable Multimodal Models. arXiv. Note: arXiv:2312.11805 [cs] External Links: Link, Document Cited by: Appendix S3, §3.1. Gifford et al. (2024) A. T. Gifford, D. Bersch, M. St-Laurent, B. Pinsard, J. Boyle, L. Bellec, A. Oliva, G. Roig, and R. M. Cichy The algonauts project 2025 challenge: how the human brain makes sense of multimodal movies. arXiv preprint arXiv:2501.00504. Cited by: §2. Harris et al. (2020) C. R. Harris, K. J. Millman, S. J. Van Der Walt, R. Gommers, P. Virtanen, D. Cournapeau, E. Wieser, J. Taylor, S. Berg, N. J. Smith, et al. Array programming with numpy. nature 585 (7825), p. 357â362. Cited by: Appendix S3. Huettel et al. (2004) S. A. Huettel, A. W. Song, and G. McCarthy Functional magnetic resonance imaging. Sinauer Associates. Cited by: §1. Hunter (2007) J. D. Hunter Matplotlib: a 2d graphics environment. Computing in Science & Engineering 9 (3), p. 90â95. External Links: Document Cited by: Appendix S3. Huth et al. (2016) A. G. Huth, W. A. de Heer, T. L. Griffiths, F. E. Theunissen, and J. L. Gallant Natural speech reveals the semantic maps that tile human cerebral cortex. Nature 532 (7600), p. 453â458 (en). Note: Number: 7600 External Links: ISSN 1476-4687, Link, Document Cited by: §1, §1, §3.2. Jalouzot et al. (2025) L. Jalouzot, A. Thual, Y. Lakretz, C. Pallier, and B. Thirion Optimizing fMRI Data Acquisition for Decoding Natural Speech with Limited Participants. arXiv (en). Note: arXiv:2505.21304 [q-bio] External Links: Link, Document Cited by: §1, §2, §4.1. Kanwisher et al. (1997) N. Kanwisher, J. McDermott, and M. M. Chun The fusiform face area: a module in human extrastriate cortex specialized for face perception. Journal of neuroscience 17 (11), p. 4302â4311. External Links: Link Cited by: §1. Kaplan et al. (2020) J. Kaplan, S. McCandlish, T. Henighan, T. B. Brown, B. Chess, R. Child, S. Gray, A. Radford, J. Wu, and D. Amodei Scaling Laws for Neural Language Models. arXiv. Note: arXiv:2001.08361 [cs] External Links: Link, Document Cited by: §4.2. Khare et al. (2025) Y. Khare, T. Peyash, A. Vanzo, and T. Yoshioka Universal-2-tf: robust all-neural text formatting for asr. arXiv preprint arXiv:2501.05948. Cited by: §1. Kingma and Ba (2017) D. P. Kingma and J. Ba Adam: A Method for Stochastic Optimization. arXiv. Note: arXiv:1412.6980 [cs] External Links: Link, Document Cited by: §4.1. LeBel et al. (2023) A. LeBel, L. Wagner, S. Jain, A. Adhikari-Desai, B. Gupta, A. Morgenthal, J. Tang, L. Xu, and A. G. Huth A natural language fMRI dataset for voxelwise encoding models. Scientific Data 10 (1), p. 555 (en). External Links: ISSN 2052-4463, Link, Document Cited by: §1, Table 1, Table 1, §2. Lerner et al. (2011) Y. Lerner, C. J. Honey, L. J. Silbert, and U. Hasson Topographic mapping of a hierarchy of temporal receptive windows using a narrated story. Journal of neuroscience 31 (8), p. 2906â2915. Cited by: §1. Lhoest et al. (2021) Q. Lhoest, A. V. Del Moral, Y. Jernite, A. Thakur, P. Von Platen, S. Patil, J. Chaumond, M. Drame, J. Plu, L. Tunstall, et al. Datasets: a community library for natural language processing. In Proceedings of the 2021 conference on empirical methods in natural language processing: system demonstrations, p. 175â184. Cited by: Appendix S3. Li et al. (2022) J. Li, S. Bhattasali, S. Zhang, B. Franzluebbers, W. Luh, R. N. Spreng, J. R. Brennan, Y. Yang, C. Pallier, and J. Hale Le petit prince multilingual naturalistic fmri corpus. Scientific Data 9 (1), p. 530. External Links: Document, ISBN 2052-4463, Link Cited by: Table 1, §2. Loshchilov and Hutter (2017) I. Loshchilov and F. Hutter Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101. Cited by: §4.1. McKinney (2010) W. McKinney Data structures for statistical computing in python.. In Proceedings of the 9th Python in Science Conference, p. 56â61. External Links: Document, Link Cited by: Appendix S3. Mesulam (1998) M. Mesulam From sensation to cognition.. Brain: a journal of neurology 121 (6), p. 1013â1052. Cited by: §1. Moeller et al. (2010) S. Moeller, E. Yacoub, C. A. Olman, E. Auerbach, J. Strupp, N. Harel, and K. UÄurbil Multiband multislice GE-EPI at 7 tesla, with 16-fold acceleration using partial parallel imaging with application to high spatial and temporal whole-brain fMRI. Magnetic Resonance in Medicine 63 (5), p. 1144â1153 (eng). External Links: ISSN 1522-2594, Document Cited by: §S1.3. Muennighoff et al. (2023) N. Muennighoff, N. Tazi, L. Magne, and N. Reimers Mteb: massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, p. 2014â2037. Cited by: §3.3. Nastase et al. (2021) S. A. Nastase, Y. Liu, H. Hillman, A. Zadbood, L. Hasenfratz, N. Keshavarzian, J. Chen, C. J. Honey, Y. Yeshurun, M. Regev, M. Nguyen, C. H. C. Chang, C. Baldassano, O. Lositsky, E. Simony, M. A. Chow, Y. C. Leong, P. P. Brooks, E. Micciche, G. Choe, A. Goldstein, T. Vanderwal, Y. O. Halchenko, K. A. Norman, and U. Hasson The âNarrativesâ fMRI dataset for evaluating models of naturalistic language comprehension. Scientific Data 8 (1), p. 250 (en). External Links: ISSN 2052-4463, Link, Document Cited by: §1, Table 1, §2. Ozcelik and VanRullen (2023) F. Ozcelik and R. VanRullen Natural scene reconstruction from fMRI signals using generative latent diffusion. arXiv. Note: arXiv:2303.05334 [cs, q-bio] External Links: Link, Document Cited by: §2. Ăzdogan et al. (2025) M. Ăzdogan, G. Landau, G. Elvers, D. Jayalath, P. Somaiya, F. Mantegna, M. Woolrich, and O. P. Jones Libribrain: over 50 hours of within-subject meg to improve speech decoding methods at scale. arXiv preprint arXiv:2506.02098. Cited by: Table 1, §2. Pasquiou et al. (2023) A. Pasquiou, Y. Lakretz, B. Thirion, and C. Pallier Information-restricted neural language models reveal different brain regionsâ sensitivity to semantics, syntax, and context. Neurobiology of Language 4 (4), p. 611â636. Cited by: §1. Paszke et al. (2019) A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, et al. Pytorch: an imperative style, high-performance deep learning library. Advances in neural information processing systems 32. Cited by: Appendix S3. Pedregosa et al. (2011) F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss, V. Dubourg, J. Vanderplas, A. Passos, D. Cournapeau, M. Brucher, M. Perrot, and E. Duchesnay Scikit-learn: Machine Learning in Python. Journal of Machine Learning Research 12 (85), p. 2825â2830. External Links: ISSN 1533-7928, Link Cited by: Appendix S3. Radford et al. (2021) A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, G. Krueger, and I. Sutskever Learning Transferable Visual Models From Natural Language Supervision. arXiv. Note: arXiv:2103.00020 [cs] External Links: Link, Document Cited by: §4.1. Reimers and Gurevych (2019) N. Reimers and I. Gurevych Sentence-bert: sentence embeddings using siamese bert-networks. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP), p. 3982â3992. Cited by: Appendix S3. Ren et al. (2025) J. Ren, N. An, C. Lin, Y. Zhang, Z. Sun, W. Zhang, S. Li, N. Guo, W. Cui, Q. Hu, W. Wang, X. Wu, Y. Wang, T. Jiang, T. D. Satterthwaite, D. Wang, and H. Liu DeepPrep: an accelerated, scalable and robust pipeline for neuroimaging preprocessing empowered by deep learning. Nature Methods 22 (3), p. 473â476 (en). External Links: ISSN 1548-7105, Link, Document Cited by: Appendix S3, §3.2. Scotti et al. (2023) P. S. Scotti, A. Banerjee, J. Goode, S. Shabalin, A. Nguyen, E. Cohen, A. J. Dempster, N. Verlinde, E. Yundler, D. Weisberg, K. A. Norman, and T. M. Abraham Reconstructing the Mindâs Eye: fMRI-to-Image with Contrastive Learning and Diffusion Priors. arXiv. Note: arXiv:2305.18274 [cs, q-bio] External Links: Link Cited by: §1, §2, §4.1. Tang et al. (2023) J. Tang, A. LeBel, S. Jain, and A. G. Huth Semantic reconstruction of continuous language from non-invasive brain recordings. Nature Neuroscience 26 (5), p. 858â866 (en). Note: Number: 5 External Links: ISSN 1546-1726, Link, Document Cited by: §1. The joblib developers (2025) The joblib developers joblib. External Links: Document, Link Cited by: Appendix S3. Toneva and Wehbe (2019) M. Toneva and L. Wehbe Interpreting and improving natural-language processing (in machines) with natural language-processing (in the brain). Advances in neural information processing systems 32. Cited by: §1. Virtanen et al. (2020) P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser, J. Bright, et al. SciPy 1.0: fundamental algorithms for scientific computing in python. Nature methods 17 (3), p. 261â272. Cited by: Appendix S3. Waskom (2021) M. L. Waskom Seaborn: statistical data visualization. Journal of Open Source Software 6 (60), p. 3021. External Links: Document, Link Cited by: Appendix S3. Wen et al. (2017) H. Wen, J. Shi, Y. Zhang, K. Lu, J. Cao, and Z. Liu Data for Neural Encoding and Decoding with Deep Learning for Dynamic Natural Vision Tests. External Links: Link, Document Cited by: §2. Wernicke (1874) C. Wernicke Der aphasische symptomencomplex: eine psychologische studie auf anatomischer basis. Max Cohn & Weigert, Breslau (de). Cited by: §1. Wolf et al. (2020) T. Wolf, L. Debut, V. Sanh, J. Chaumond, C. Delangue, A. Moi, P. Cistac, T. Rault, R. Louf, M. Funtowicz, et al. Transformers: state-of-the-art natural language processing. In Proceedings of the 2020 conference on empirical methods in natural language processing: system demonstrations, p. 38â45. Cited by: Appendix S3. Xiao et al. (2025) C. Xiao, H. P. Chan, H. Zhang, W. Xu, M. Aljunied, and Y. Rong Scaling language-centric omnimodal representation learning. Advances in Neural Information Processing Systems. Cited by: Appendix S3, §3.3. Yang et al. (2022) Y. Yang, M. Hira, Z. Ni, A. Astafurov, C. Chen, C. Puhrsch, D. Pollack, D. Genzel, D. Greenberg, E. Z. Yang, et al. Torchaudio: building blocks for audio and speech processing. In ICASSP 2022-2022 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), p. 6982â6986. External Links: Document, Link Cited by: Appendix S3. Zhang et al. (2025) Y. Zhang, M. Li, D. Long, X. Zhang, H. Lin, B. Yang, P. Xie, A. Yang, D. Liu, J. Lin, F. Huang, and J. Zhou Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models. arXiv. Note: arXiv:2506.05176 [cs] External Links: Link, Document Cited by: Appendix S3, §1, §3.3. Appendix S1 A visual glimpse at the data Figure S1 shows a single brain data frame (or TR) from Cephalonauts One, after preprocessing and before masking. To the expert as well as to the untrained eye, this shows that brain recordings are typically noisy, and may contain reconstruction artifacts. However, as demonstrated in the decoding results of Section 4.2, these data still contain meaningful information, and decoders can successfully learn to leverage it to predict precise information about the perceptual experience of the participants. Figure S1: Single brain data frame from Cephalonauts One. This image is a screenshot obtained with niivue (https://niivue.com/) of a single brain volume from Cephalonauts One, after preprocessing and before masking. It shows all three views of the brain (axial, coronal and sagittal) in the native anatomical space of one participant. S1.1 Interactive exploration notebook To make the dataset easier to inspect, we provide an interactive Google Colab notebook: colab.research.google.com/drive/1YTF_yOmYMuM8Fg69BTjRaGwIQvrBU3U1. The notebook is a lightweight companion to the Hugging Face repository and the baseline code. It guides users through authenticating to Hugging Face, loading one validation Parquet file with the datasets library, inspecting the available fields, and visualizing example brain and stimulus data. The example corresponds to the validation split for subject 1 in the fsaverage5 preprocessing space, and includes the aligned fMRI vector, text and audio embeddings, transcript context, session and run identifiers, TR index, timing offsets, and source audio and fMRI paths. It is intended for quick exploration in a browser, without requiring a local installation. Full-scale preprocessing, training, and evaluation should still be run from the released code repository. S1.2 Participant instructions and comprehension questions At the beginning of each run, participants received the following instruction: Vous allez entendre un texte. Ăcoutez-le attentivement, sans bouger. Vous pouvez fermer les yeux si vous le souhaitez. which translates to: You will hear a text. Listen to it attentively, without moving. You can close your eyes if you wish. At the end of each run, participants answered three multiple-choice comprehension questions. For example: Comment les LLM comme Claude gĂŠnèrent-ils du texte ? A. En analysant le sens profond des mots grâce Ă une conscience artificielle. B. En prĂŠdisant token par token (mots, ponctuation) via des calculs de probabilitĂŠs statistiques. C. En recopiant des phrases existantes stockĂŠes dans une base de donnĂŠes gĂŠante. This example question refers to a podcast segment about large language models. It translates to: How do LLMs such as Claude generate text? A. By analyzing the deep meaning of words through artificial consciousness. B. By predicting tokens one by one (words, punctuation) through statistical probability computations. C. By copying existing sentences stored in a giant database. The correct answer is B. S1.3 MRI scanner parameters All images were acquired on a 3.0T Siemens MAGNETOM Cima.X whole-body MRI scanner using a 64-channel head/neck coil. Core acquisition parameters for the functional and anatomical scans are summarized in Table S1 while sequence-specific details are provided below. Table S1: Core acquisition parameters for the functional and anatomical MRI sequences. Parameter Functional MRI Anatomical MRI Sequence family 2D gradient-echo EPI 3D T1-weighted spoiled gradient-echo Orientation Transverse Sagittal FOV (read Ă phase) 198Ă198198Ă 198 m2 230Ă230230Ă 230 m2 Voxel size 3.0Ă3.0Ă3.03.0Ă 3.0Ă 3.0 m3 0.8Ă0.8Ă0.80.8Ă 0.8Ă 0.8 m3 Slice thickness 3.0 m 0.8 m TR 520 ms 6.20 ms Flip angle 46â46 9â9 Bandwidth 2444 Hz/pixel 480 Hz/pixel Acquisition time 15 min 10 s 3 min 28 s Functional MRI acquisition Functional images were acquired using a 2D gradient-echo echo-planar imaging (EPI) sequence [Moeller et al., 2010] sensitive to blood-oxygen-level-dependent (BOLD) contrast. The acquisition comprised 48 transverse slices with no slice gap, acquired in interleaved order, with an acquisition matrix of 66Ă6666Ă 66. Sequence parameters included TE =30.0=30.0 ms, posterior-to-anterior phase encoding, and fat saturation. Slice acceleration used simultaneous multi-slice imaging with factor 6, with no in-plane parallel imaging. Other notable parameters were echo spacing =0.54=0.54 ms and phase-scrambling was enabled. Each run contained 1736 multi-band volumes together with one single-band reference volume. Anatomical MRI acquisition High-resolution anatomical images were acquired using a 3D T1-weighted spoiled gradient-echo Dixon sequence with non-selective excitation. The acquisition consisted of one sagittal slab with 230 slices, reconstructed at isotropic resolution. The readout was Cartesian with bipolar acquisition. The base resolution was 288, with phase resolution =100%=100\% and slice resolution =100%=100\%. Two echoes were acquired for waterâfat separation (TE1=12.46_1=12.46 ms, TE2=23.95_2=23.95 ms), and images were reconstructed as magnitude volumes. Parallel imaging used a total acceleration factor of 4 (2Ă22Ă 2, phase-encoding Ă partition), with 24 reference lines in each accelerated direction. No partial Fourier or elliptical scanning was used. Figure S2: Decoding framework. A. Brain activity is represented as a sequence of fMRI volumes sampled every 520 ms (=TR). For each frame t, we associate a stimulus t z_t derived from the content perceived while the brain data was recorded during that segment, using either the raw audio or the transcript. Then, a stimulus embedding t y_t is computed over a window of Ď consecutive segments (stimulus context). The brain data x used to decode t y_t consists of brain volumes t w_t temporally smoothed over a window of Îş frames, and taken with a time lag δ with regard to t. In this illustration, the temporal smoothing window Îş is set to 3 frames â or TRs, i.e. it is set to 3 Ă 520ms â the stimulus context Ď is set to 3 frames, and the lag δ is set to 2 frames. B. A pretrained encoder produces stimulus embeddings Y from the stimulus Z, and a decoder fθf_θ is trained to predict these embeddings from the brain recordings X. Decoding can therefore be performed toward text-based or audio-based stimulus representations or even a concatenation of both. In this figure, variables used to decode stimuli perceived during frame t are shown in black, while variables not used for decoding are shown with reduced opacity. S1.4 Formal definition of the brain decoding problem For a given subject pâPpâ P, a given session sâSsâ S and a given run râRrâ R in that session, we denote by p,s,rââTs,rĂV W_p,s,r ^T_s,rĂ V the matrix of recorded brain activity during that run, where Ts,rT_s,r is the number of brain volumes acquired during that run and V the number of voxels. Let s,rââTs,rĂD Y_s,r ^T_s,rĂ D be the matrix of stimulus features derived from the podcast â either its transcript or the raw audio â listened to during that run, where D is the dimensionality of the stimulus features. Multiple types of stimulus features s,r Y_s,r are available, as described in Section 3.3. Let ptrain W_p^train be the concatenation of all brain data recorded for subject p across all sessions and runs that are used for training. We define ptrain Y_p^train similarly for the stimulus features, and pval W_p^val, ptest W_p^test, pval Y_p^val, ptest Y_p^test for the validation and test sets. The set of training, validation and test sessions are disjoint. ptrainââ¨sâStrainrâRp,s,rââTtrainĂV, W_p^train\; \; _ subarraycsâ S_train\\ râ R subarray W_p,s,r ^T_trainĂ V, ptrainââ¨sâStrainrâRp,s,rââTtrainĂD, Y_p^train\; \; _ subarraycsâ S_train\\ râ R subarray Y_p,s,r ^T_trainĂ D, SâStrainâSvalâStestS\; \;S_train\; \;S_val\; \;S_test Let ptrain X_p^train be the preprocessed brain data obtained from ptrain W_p^train after applying the preprocessing pipeline described in Section 3.2, and after applying a temporal smoothing function h with window size Îş, such that tââV x_t ^V is computed from the Îş brain volumes w centered on t, as illustrated in Figure S2. The decoding problem consists in learning a function fθf_θ parameterized by θ that maps brain activity to stimulus features, and in evaluating the performance of this function on held-out test data. We refer to fθf_θ as the decoder. Since we only train models on data from a single subject, we drop the subscript p for readability. For a given brain volume tâTtâ T, the model takes as input the vector of brain activity tââV x_t ^V and produces as output a vector of predicted stimulus features ^tâfθâ(t)ââD y_t f_θ( x_t) ^D. Using gradient descent, θ is optimised to minimise a loss function âL that quantifies the discrepancy between the predicted embeddings ^t y_t and the true embeddings t y_t. A temporal lag δ can also be introduced, to account for the delay of the hemodynamic response for instance. The full model can be written as: ^tâfθâ(t+δ) y_t\; \;f_θ\,( x_t+δ) Additionally, the model can be trained on multiple consecutive brain volumes x, and tasked with predicting stimulus features derived from a window of multiple consecutive stimulus segments, which we did not explore in the present work but which constitute interesting avenues for future research using Cephalonauts One. Appendix S2 Preprocessing and decoder training details Computational resources Preprocessing and decoder training were performed on H100-80G GPU workers. Across all preprocessing and experiments, including preliminary runs not reported in the paper, the total compute was approximately 660 H100-80G GPU-hours. About 400 GPU-hours were run on cloud computing resources, and about 260 GPU-hours used HPC resources from GENCIâIDRIS (Grant 2025-AD010316590R1). Most of this compute was used for fMRI preprocessing with DeepPrep. The baseline decoder itself can also run on consumer hardware, requiring less than 8 GB of GPU memory and less than 8 GB of system RAM. Decoder hyperparameters Table S2 reports the baseline decoder settings used across the decoding experiments. These settings define the input fMRI window, hemodynamic lag, stimulus context windows, embedding models, optimization procedure, and retrieval evaluation protocol. Table S2: Decoding hyperparameters. Hyperparameters used for all decoding experiments unless otherwise specified. Hyperparameter Value Decoder architecture one-hidden-layer MLP Hidden dimension 4096 Activation function GELU Dropout 0.5 Hemodynamic lag 2.5 seconds Text context window 20 words Audio context window 10 seconds Text embedding model Qwen3-Embedding-8B Audio embedding model LCO-Embedding-Omni-7B Optimizer AdamW Learning rate 3e-4 Weight decay 3e-4 Batch size 2048 Number of epochs 100 max Loss function CLIP-style contrastive loss Temperature 2e-2 Early stopping patience 5 Validation criterion Median Relative Rank Retrieval set size âź 2000 Hardware NVIDIA H100-80G GPU workers Figure S3: Supplementary scaling analysis across brain spaces and stimulus targets. Median Relative Rank is shown as a function of subject-specific training hours for each released brain representation and stimulus target. The brain representations correspond to the preprocessing outputs listed in Table 3: fsaverage5, T1w, and MNI. The stimulus targets correspond to the text, audio, and concatenated embeddings described in Table 4. Lower values indicate better retrieval performance. Scaling analysis Figure S3 extends the main scaling analysis to all released brain preprocessing spaces and stimulus embedding targets. All combinations provide useful decoding signal: Median Relative Rank improves with additional training data for the fsaverage5, T1w, and MNI representations, and for text, audio, and concatenated embedding targets. This supports the relevance of the preprocessing outputs released with the dataset, described in Table 3, and of the stimulus embeddings described in Table 4. Appendix S3 Licenses and terms of use for existing assets Code dependencies The released code uses Python, NumPy [Harris et al., 2020], SciPy [Virtanen et al., 2020], PyTorch [Paszke et al., 2019], torchaudio [Yang et al., 2022], Hugging Face datasets [Lhoest et al., 2021], Hugging Face transformers [Wolf et al., 2020], Hugging Face Hub tools, scikit-learn [Pedregosa et al., 2011], nilearn [Abraham et al., 2014], sentence-transformers [Reimers and Gurevych, 2019], pandas [McKinney, 2010], matplotlib [Hunter, 2007], seaborn [Waskom, 2021], joblib [The joblib developers, 2025], tqdm [da Costa-Luis et al., 2026], uv, and pip. Embedding models and hosted services Transcript embeddings were computed with Qwen3-Embedding-8B [Zhang et al., 2025], released under the Apache License 2.0. Audio embeddings were computed with LCO-Embedding-Omni-7B [Xiao et al., 2025], also released under the Apache License 2.0. Transcripts were generated with AssemblyAI Universal 3 Pro; this use was governed by the AssemblyAI Terms of Service.22 2 https://w.assemblyai.com/legal/terms-of-service Comprehension questions were generated with Gemini 3 Pro [Gemini Team and others, 2024]; this use was governed by the Google APIs Terms of Service and the Gemini API Additional Terms of Service.33 3 https://ai.google.dev/gemini-api/terms No Gemini model weights are redistributed as part of this work. Preprocessing software The fMRI preprocessing used DeepPrep [Ren et al., 2025], which is released under the Apache License 2.0. DeepPrep uses standard neuroimaging templates distributed through TemplateFlow [Ciric et al., 2022]. DeepPrep relies on FreeSurfer components for anatomical processing [Fischl, 2012]; FreeSurfer is distributed under its own software license and requires a free license file for use.44 4 https://freesurfer.net/fswiki/License NeurIPS Paper Checklist 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paperâs contributions and scope? Answer: [Yes] Justification: The abstract and Introduction state that the paper contributes a deep fMRI dataset, paired audio/transcript representations, a retrieval benchmark, baseline decoders, and a scaling analysis. The Discussion and Limitations sections describe the main scope restrictions, including the small number of subjects, single-site acquisition, French podcast stimuli, and retrieval-based evaluation. Guidelines: ⢠The answer [N/A] means that the abstract and introduction do not include the claims made in the paper. ⢠The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers. ⢠The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. ⢠It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. 2. Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: The Discussion contains a dedicated Limitations subsection. It discusses the limited number of subjects, single-site acquisition, restricted language and stimulus domain, and retrieval-based evaluation. Guidelines: ⢠The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper. ⢠The authors are encouraged to create a separate âLimitationsâ section in their paper. ⢠The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. ⢠The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. ⢠The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. ⢠The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. ⢠If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. ⢠While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that arenât acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. 3. Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [N/A] Justification: The paper introduces a dataset and empirical decoding benchmark. It does not present theoretical results, theorems, or proofs. Guidelines: ⢠The answer [N/A] means that the paper does not include theoretical results. ⢠All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. ⢠All assumptions should be clearly stated or referenced in the statement of any theorems. ⢠The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. ⢠Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. ⢠Theorems and Lemmas that the proof relies upon should be properly referenced. 4. Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: The paper describes the released data representations, stimulus preprocessing, fMRI alignment, train/validation/test separation, retrieval metrics, baseline decoder, optimizer, loss, and hyperparameters. The data and code release is linked in the abstract and is intended to provide the scripts needed to reproduce the main decoding results. Guidelines: ⢠The answer [N/A] means that the paper does not include experiments. ⢠If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. ⢠If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. ⢠Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. ⢠While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The abstract provides a public release link for the dataset and code. The release is intended to include the preprocessed fMRI data, stimulus representations, benchmark splits, metrics, and baseline code needed to reproduce the main results. Guidelines: ⢠The answer [N/A] means that paper does not include experiments requiring code. ⢠Please see the NeurIPS code and data submission guidelines (https://neurips.c/public/guides/CodeSubmissionPolicy) for more details. ⢠While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). ⢠The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://neurips.c/public/guides/CodeSubmissionPolicy) for more details. ⢠The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. ⢠The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. ⢠At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). ⢠Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. 6. Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results? Answer: [Yes] Justification: The Decoding experiments section specifies the held-out-session retrieval task, the use of all segments from the held-out session as retrieval candidates, the metrics, baseline decoder, optimizer, and loss. Appendix S2 reports the hyperparameters used for the baseline experiments. Guidelines: ⢠The answer [N/A] means that the paper does not include experiments. ⢠The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. ⢠The full details can be provided either with the code, in appendix, or as supplemental material. 7. Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: Results are reported with error bars across five folds, using session indices 19â23 once each for test. For each fold, the next session in this set is held out for validation (wrapping from 23 to 19), and the remaining 22 sessions form the training pool. The error bars in the scaling-law plots are empirical one-standard-deviation error bars computed by seaborn with errorbar="sd" across the fold results. They therefore capture variability due to the held-out session split, not model initialization. No normality assumption or formal significance test is used; the bars are standard deviations, not standard errors or confidence intervals. Guidelines: ⢠The answer [N/A] means that the paper does not include experiments. ⢠The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. ⢠The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). ⢠The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) ⢠The assumptions made should be given (e.g., Normally distributed errors). ⢠It should be clear whether the error bar is the standard deviation or the standard error of the mean. ⢠It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. ⢠For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates). ⢠If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The appendix paragraph on computational resources reports the H100-80G GPU workers used for preprocessing and decoder training. Guidelines: ⢠The answer [N/A] means that the paper does not include experiments. ⢠The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. ⢠The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. ⢠The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didnât make it into the paper). 9. Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.c/public/EthicsGuidelines? Answer: [Yes] Justification: The research conducted in this paper conforms, in every respect, with the NeurIPS Code of Ethics. Guidelines: ⢠The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics. ⢠If the authors answer [No] , they should explain the special circumstances that require a deviation from the Code of Ethics. ⢠The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). 10. Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: The Broader impact section covers positive uses in neuroscience and multimodal representation learning, as well as potential risks such as privacy concerns and misuse of brain decoding results. It also describes the gated access terms used for the dataset release. Guidelines: ⢠The answer [N/A] means that there is no societal impact of the work performed. ⢠If the authors answer [N/A] or [No] , they should explain why their work has no societal impact or why the paper does not address societal impact. ⢠Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. ⢠The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. ⢠The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. ⢠If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11. Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)? Answer: [Yes] Justification: The Broader impact section states that brain data are sensitive biomedical data and describes the safeguards used for the dataset release. Access to Cephalonauts One is gated on Hugging Face, requires acceptance of terms of use, prohibits attempts to identify or contact study volunteers, requires compliance with applicable institutional and ethics-review rules, and requires redistribution under the same conditions. Guidelines: ⢠The answer [N/A] means that the paper poses no such risks. ⢠Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. ⢠Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. ⢠We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. 12. Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: Appendix S3 credits the main code dependencies, embedding models, hosted Gemini service, and preprocessing software used in the work. It also lists applicable licenses or terms of use for the pretrained models and preprocessing tools. Guidelines: ⢠The answer [N/A] means that the paper does not use existing assets. ⢠The authors should cite the original paper that produced the code package or dataset. ⢠The authors should state which version of the asset is used and, if possible, include a URL. ⢠The name of the license (e.g., C-BY 4.0) should be included for each asset. ⢠For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. ⢠If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. ⢠For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. ⢠If this information is not available online, the authors are encouraged to reach out to the assetâs creators. 13. New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: The paper documents the new dataset, released brain and stimulus representations, benchmark task, metrics, baseline model, and hyperparameters. The dataset and code release is intended to provide accompanying documentation for accessing and using these assets. Guidelines: ⢠The answer [N/A] means that the paper does not release new assets. ⢠Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. ⢠The paper should discuss whether and how consent was obtained from people whose asset is used. ⢠At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. 14. Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [Yes] Justification: Appendix S1.2 provides the full instruction text given at the beginning of each run and an example multiple-choice comprehension question. The Dataset section reports monetary compensation; screenshots are not applicable to this fMRI listening protocol. Guidelines: ⢠The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. ⢠Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. ⢠According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. 15. Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? Answer: [Yes] Justification: The Dataset section states that participants were healthy adult volunteers who met standard MRI safety criteria, that the study was approved by an ethics committee (CPP Ile de France 2 2025-A02302-47 on February 16, 2026), that all participants gave written informed consent, and that participants received monetary compensation. Guidelines: ⢠The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. ⢠Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. ⢠We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. ⢠For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. 16. Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required. Answer: [Yes] Justification: The paper states that Gemini 3 Pro was used to generate comprehension questions, transcript segments were embedded with Qwen3-Embedding-8B, and audio segments were embedded with LCO-Embedding-Omni-7B. These uses are described in the Dataset and Stimulus sections. Guidelines: ⢠The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. ⢠Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.