MEG-to-MEG Transfer Learning and Cross-Task Speech/Silence Detection with Limited Data
Authors: Xabier de Zuazo, Vincenzo Verbeni, Eva Navas, Ibon Saratxaga, Mathieu Bourguignon, Nicola Molinaro
Organizations: HiTZ Center, University of the Basque Country – UPV/EHU, Spain · Basque Center on Cognition, Brain and Language – BCBL, Spain · Laboratory of Functional Anatomy, Faculty of Human Motor Sciences, Université libre de Bruxelles (ULB), Belgium · Laboratoire de Neuroanatomie et Neuroimagerie translationnelles (LN2T), ULB Neuroscience Institute, Université libre de Bruxelles (ULB), Belgium · WEL Research Institute, Belgium · Ikerbasque, Basque Foundation for Science, Spain
Data-efficient neural decoding is a central challenge for speech brain-computer interfaces. We present the first demonstration of transfer learning and cross-task decoding for MEG-based speech models spanning perception and production. We pre-train a Conformer-based model on 50 hours of single-subject listening data and fine-tune on just 5 minutes per subject across 18 participants. Transfer learning yields consistent improvements, with in-task accuracy gains of 1-4% and larger cross-task gains of up to 5-6%. Not only does pre-training improve performance within each task, but it also enables reliable cross-task decoding between perception and production. Critically, models trained on speech production decode passive listening above chance, confirming that learned representations reflect shared neural processes rather than task-specific motor activity.
Figures & tables
Task
Accu. (%)
F1 (%)
AUC (%)
Scratch (baseline)
listen
76.2 ± 4.8
85.5 ± 3.2
64.0 ± 9.4
playback
75.1 ± 6.1
84.0 ± 4.7
67.7 ± 4.2
production
83.6 ± 5.4
89.7 ± 3.5
81.1 ± 8.6
Transfer Learning
listen
79.0 ± 4.8
87.7 ± 3.2
68.7 ± 6.2
Table 1: In-task evaluation results (mean ± std across subjects, in %). Improvements over baseline are marked in bold.
Train → Test
Accu. (%)
F1 (%)
AUC (%)
listen to play.
72.5 ± 3.6
83.0 ± 2.7
59.4 ± 5.3
listen to prod.
71.5 ± 3.5
82.3 ± 2.6
57.9 ± 7.4
play. to listen
73.4 ± 2.9
83.7 ± 2.2
59.4 ± 5.7
play. to prod.
71.4 ± 4.4
82.1 ± 3.6
57.4 ± 6.7
prod. to listen
66.1 ± 6.3
77.7 ± 5.8
54.0 ± 6.3
prod. to play.
65.0 ± 5.9
76.5 ± 5.5
54.6 ± 4.7
Table 2: Cross-task decoding results for the model trained from scratch (mean ± std across subjects, in %). All reported cross-task results are significantly above chance.
Train → Test
Accu. (%)
F1 (%)
AUC (%)
listen to play.
76.9 ± 4.0
86.5 ± 2.8
61.1 ± 4.5
listen to prod.
75.3 ± 5.2
85.3 ± 3.9
57.3 ± 7.3
play. to listen
78.0 ± 3.9
87.1 ± 2.9
61.6 ± 5.8
play. to prod.
73.6 ± 7.1
83.7 ± 6.1
58.1 ± 9.1
prod. to listen
69.3 ± 8.0
80.1 ± 7.3
56.0 ± 5.7
prod. to play.
68.3 ± 7.4
79.0 ± 7.2
56.6 ± 3.8
Table 3: Cross-task decoding results for the pretrained model (mean ± std across subjects, in %). Improvements over baseline are marked in bold.
Figure 1: Effect size of transfer learning for cross-task decoding. Bars show the mean improvement in F1 score (transfer learning minus training from scratch) for each task and train-test task pair. Error bars indicate the standard deviation across subjects.
Figure 2: Cross-task F1 asymmetry for each task pair. Each x-axis category shows two directions: circle markers correspond to the left-to-right direction in the label, and diamond markers to the reverse direction. For each model, the two direction-specific points are connected to highlight asymmetry. Error bars indicate the standard deviation across subjects.
Figure 3: Subject-level in-task F1 improvement with transfer learning. Each bar shows the improvement in F1 score (transfer learning minus training from scratch) for a single subject and task. Colors indicate the three tasks.
We propose MEBM-Speech, a multi-scale enhanced neural decoder for speech activity detection from non-invasive magnetoencephalography (MEG) signals. Built upon the BrainMagic backbone, MEBM-Speech integrates three complementary temporal modeling mechanisms: a multi-scale convolutional module for short-term pattern extraction, a bidirectional LSTM (BiLSTM) for long-range context modeling, and a depthwise separable convolutional layer for efficient cross-scale feature fusion. A lightweight temporal jittering strategy and average pooling further improve onset robustness and boundary stability. The model performs continuous probabilistic decoding of MEG signals, enabling fine-grained detection of speech versus silence states - an ability crucial for both cognitive neuroscience and clinical applications. Comprehensive evaluations on the LibriBrain Competition 2025 Track1 benchmark demonstrate strong performance, achieving an average F1 macro of 89.3% on the validation set and comparable results on the official test leaderboard. These findings highlight the effectiveness of multi-scale temporal representation learning for robust MEG-based speech decoding.
Li Songyi, Zheng Linze, Liang Jinghua +1
Speech and Hearing Research Center, School of Intelligence Science and Technology, Peking University, Beijing, China 100871 · Center for BioMed-X Research, Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China 100871
EEG foundation models pretrained on thousands of hours have shown large gains over task-specific networks for motor imagery, seizure detection, sleep staging, and emotion recognition, but their transfer to speech decoding - arguably the most demanding non-invasive BCI application - remains untested. We present the first systematic benchmark of EEG foundation models against strong convolutional baselines for speech decoding, using two corpora: UGR-MINDVOICE (overt and covert Iberian Spanish) and BCI Competition 2020 Track 3 (imagined speech). We compare two foundation models (LaBraM, EEGMamba) against three established baselines (EEGNet, ShallowFBCSPNet, EEGConformer) under a unified preprocessing and fine-tuning protocol. Large-scale EEG pretraining yields no consistent advantage over a 16K-parameter CNN on speech tasks, indicating that current general-purpose EEG pretraining does not yet transfer to speech production and motivating speech-specific foundation models.
Owais Mujtaba Khanday, Mohamed Baha Ben Ticha, Sanae Belfrouh +2
Dept. Signal Theory, Telematics and Communications, University of Granada, Spain · Research Centre for Information and Communication Technologies (CITIC-UGR), Spain · Laboratory of Information Technologies, University of Chouaib Doukkali, Morocco +1
Decoding speech from non-invasive brain signals is challenging. For the LibriBrain 2025 Speech Detection task, we propose a novel two-step framework that bypasses direct reconstruction. First, a contrastive learning model retrieves the matching speech segment for the given test MEG from a large-scale audio library (LibriVox). Second, a speech detection model generates the binary silence/speech sequence directly from this retrieved audio. With this approach, our team Sherlock Holmes achieved first place in the extended track (F1-score: 0.962), demonstrating that leveraging external audio databases is a highly effective strategy.
Boda Xiao, Bo Wang, Heping Cheng
Center for BioMed-X Research, Academy for Advanced Interdisciplinary Studies, Peking University, Beijing, China · Speech and Hearing Research Center, School of Intelligence Science and Technology, Peking University, Beijing, China · National Biomedical Imaging Center, State Key Laboratory of Membrane Biology, Institute of Molecular Medicine, Peking-Tsinghua Center for Life Sciences, College of Future Technology, Peking University, Beijing, China