- The paper introduces MEBM-Speech, a 10.3-million-parameter MEG speech activity detector that combines BrainMagic encoders, multi-scale convolutions, and BiLSTM temporal modeling with continuous probabilistic decoding.
- The model reaches 89.34% ± 0.24 F1_macro and 89.25% ± 0.37 Acc_macro on Sherlock1 validation sessions, with comparable performance on the official online test server.
- The BrainMagic encoder provides the largest ablation benefit, while multi-scale convolution and BiLSTM offer smaller gains, highlighting open questions about cross-subject generalization and real-time deployment cost.
MEBM-Speech addresses speech activity detection (SAD) from non-invasive magnetoencephalography (MEG) in the LibriBrain Competition 2025 Track 1, where the task is to segment continuous MEG recordings into speech versus silence intervals while a participant listens to natural audiobook stimuli (Landau et al., 11 Jun 2025, Özdogan et al., 2 Jun 2025). The system extends the BrainMagic backbone (2603.02255) with multi-scale temporal modeling and reframes the task as continuous probabilistic decoding rather than framewise binary classification.
The central methodological departure from the official baseline is the treatment of SAD as continuous probabilistic decoding: the model outputs a time-varying probability sequence P∈[0,1]1×T, trained with an MSE loss against binary labels, and binarized post hoc via adaptive thresholding. The authors argue this matches the smooth temporal evolution of neural responses at speech onsets and offsets. Two training-time choices reinforce boundary robustness: temporal jittering, which randomly shifts phoneme onset labels within ±2 frames (±20 ms at 100 Hz), and average pooling for temporal smoothing. Model selection retains the five checkpoints with lowest validation loss and jointly optimizes over 99 classification thresholds (0.01–0.99) against F1_macro — a selection procedure that ties threshold choice tightly to validation data, which may partially explain the small gap between local and leaderboard scores.
Architecture
The input is restricted to the 204 gradient (grad) MEG channels, downsampled to 100 Hz and normalized per 12-second window (T=1200, step size 6 s). A spatial attention module first recalibrates channel responses into a D=128-dimensional representation. This representation feeds three parallel temporal branches:
- Five BrainMagic encoders capturing mid-term contextual dependencies;
- Twelve multi-scale convolutional blocks extracting fine-grained local features across multiple receptive fields;
- A BiLSTM modeling long-range dependencies.
Branch outputs are concatenated and fused by a depthwise separable convolution, followed by average pooling (window 31, stride 15) and linear-interpolation upsampling to per-sample probabilities. The full model has approximately 10.3M parameters, converges within 10 epochs (~20 minutes on one A800 GPU), using AdamW with learning rate 10−3 and dropout 0.01.
Results
On an offline validation set constructed from Sherlock1 sessions 11–12, averaged over six random seeds, the full model achieves 89.34% ± 0.24 F1_macro and 89.25% ± 0.37 Acc_macro, with a comparable ~89% F1_macro on the official online test server, indicating generalization to held-out sessions.
| Variant |
F1_macro (%) |
Acc_macro (%) |
| Full model |
89.34 ± 0.24 |
89.25 ± 0.37 |
| w/o BM Encoder |
88.36 ± 0.13 |
88.29 ± 0.24 |
| w/o Multi-scale Conv |
89.17 ± 0.19 |
88.98 ± 0.28 |
| w/o BiLSTM |
89.21 ± 0.17 |
89.27 ± 0.22 |
| w/o BM Enc + Multi-scale Conv |
87.91 ± 0.13 |
87.76 ± 0.38 |
| w/o Multi-scale Conv + BiLSTM |
89.17 ± 0.21 |
89.20 ± 0.24 |
| w/o BiLSTM + BM Encoder |
85.59 ± 0.20 |
85.47 ± 0.18 |
The ablations indicate that the BM encoder is the dominant component: removing it costs roughly 1 point alone and nearly 4 points when combined with BiLSTM removal (85.59%). By contrast, removing only the multi-scale convolution or the BiLSTM degrades performance by well under half a point, and in one configuration (w/o Multi-scale Conv + BiLSTM) accuracy differences fall near seed-level variability. The paper's claim that all three branches contribute "complementary" features is thus supported most strongly for the BM encoder; the marginal contributions of the recurrent branch are modest relative to its added cost, a point worth weighing when considering deployment.
Limitations and open questions
Several constraints qualify these results. Validation uses Sherlock1 sessions 11–12 rather than the true holdout distribution, so the reported local scores are an approximation of competition conditions. The adaptive threshold is selected on validation data, introducing a mild selection bias. All evaluation appears within-subject on a single stimulus narrative; cross-subject generalization — acknowledged by the authors as future work — remains untested, as does real-time operation and detection during speech production. Finally, the ablation does not report computational cost per variant, leaving unclear whether the BiLSTM's sub-point gain justifies its inference latency in a BCI setting.
Conclusion
MEBM-Speech demonstrates that a compact (~10.3M parameter), multi-branch extension of BrainMagic with continuous probabilistic decoding, label jittering, and grad-only channel selection achieves ~89% F1_macro for MEG-based speech detection on the LibriBrain benchmark. The evidence most firmly establishes the value of the BrainMagic encoder and the probabilistic decoding formulation; the contributions of the auxiliary branches are smaller and their cost-effectiveness remains an open question for real-time, cross-subject speech BCIs.