MEGConformer: Conformer-Based MEG Decoder for Robust Speech and Phoneme Classification

1. Introduction

• The authors adapt Conformer, a convolution-augmented Transformer, based on its combination of self-attention and convolution for temporal and spectral MEG patterns.They describe Conformer architectures as state-of-the-art in ASR and hypothesize that their strengths suit MEG speech tasks.Source: §1 · p. 1
• The study evaluates raw-MEG windows for speech activity or phoneme identity using official data splits and evaluation metrics.The models share a backbone but differ in input and output processing.Source: §1 · p. 1
• The work contributes a shared Conformer architecture, input normalization for holdout generalization, MEG-specific augmentation, and dynamic grouping for averaged phoneme samples.The authors report that their system won the Phoneme Classification task.Source: §1 · p. 1

2. Related work

• EEG work has used CNN- and Transformer-style models for speech-feature or spectrogram reconstruction, while MEG studies include word- and phrase-level decoding and phoneme-level tasks.MEG approaches described include wavelet denoising, CNNs, transfer learning, compact end-to-end networks, and transformer-based neural encoding.Source: §2 · p. 1
• LibriBrain introduced within-subject MEG data and standardized tasks for speech detection, phoneme classification, and keyword spotting.The passage also notes contrastive and self-supervised objectives, foundation-model guidance, and brain-to-text systems with language-model rescoring.Source: §2 · p. 1
• Recent non-invasive brain-to-text systems combine discriminative decoders with language-model rescoring and cross-dataset scaling.The passage reports significant score improvements for this approach.Source: §2 · p. 2

3. Dataset

• LibriBrain contains over 50 hours of MEG from one participant listening to Sherlock Holmes audiobooks, with time-aligned speech activity and phoneme annotations.The recordings use a 306-channel Elekta/MEGIN system and are downsampled to 250 Hz following the competition protocol.Source: §3 · p. 2
• The study follows the official Standard-track train, validation, test, and holdout splits and uses F1-macro; the holdout split is not publicly available.Speech detection is binary speech-versus-silence classification, with speech occupying about 76.7% of labeled time.Source: §3 · p. 2
• Phoneme windows use 39 ARPAbet labels and have a long-tailed occurrence distribution; competition holdout examples are averages over 100 samples.The training setup mirrors the holdout averaging condition using dynamic grouping.Source: §3 · p. 2

4.1. Model

• A lightweight convolutional projection maps 306-channel MEG windows into a shared Conformer encoder, with compact task-specific model configurations.Speech uses 2.5-second windows and a 16-layer model; phoneme classification uses 0.5-second windows and a custom 7-layer model.Source: §4.1 · p. 2

4.2. Speech detection task specifics

• Speech detection uses binary cross-entropy with label smoothing and MEGAugment, which masks time spans and narrow frequency bands; training windows overlap.Inference removes predicted speech sequences shorter than 60 samples, or 240 ms.Source: §4.2 · p. 2

4.3. Phoneme classification task specifics

• Phoneme classification uses 39 output logits, cross-entropy, and inverse-square-root class weighting to address class imbalance.Instance-level normalization is applied after 100-sample averaging, normalizing each channel over that window's time axis without accumulated dataset or batch statistics.Source: §4.3 · p. 2
• A 100-sample dynamic grouping loader reshuffles groups each epoch to create independent averages of the same class while preserving temporal locality.The final phoneme predictions ensemble the five best seeds using majority voting.Source: §4.3 · p. 3

5. Experimental setup

• All experiments used the official train, validation, test, and holdout partitions; re-partitioning or rebalancing did not produce consistent improvements.The original configuration was retained throughout.Source: §5 · p. 3
• Models shared AdamW settings: learning rate 1·10−4, weight decay 5·10−2, and batch size 256, with validation F1-macro early stopping at patience 10.The final checkpoint was selected using validation F1-macro.Source: §5 · p. 3
• Evaluation used F1macro as the primary metric on the test and holdout splits.Source: §5 · p. 3
• Each configuration was trained with ten random seeds on individual NVIDIA A100 or H100 GPUs.Source: §5 · p. 3

6. Results

• 88.9% F1-macro in Speech Detection surpassed the official baseline (68.0%), while 65.8% F1-macro on Phoneme Classification holdout won the Standard track.The Standard-track phoneme result also exceeded results reported by Extended-track systems despite their use of additional data sources.Source: §6 · p. 3
• Instance-level normalization improved holdout generalization by over 200%, compared with +17.8% for batch normalization and +88.2% for layer normalization.The authors describe instance-level normalization as essential for holdout generalization.Source: §6 · p. 3
• 55.09±1.77% across five seeds before ensembling rose to 65.82% on phoneme holdout with majority voting, boosting scores by 19.48%.For Speech Detection, smoothing raised test F1macro from 88.12% to 88.37%, with a final holdout score of 88.90%.Source: §6 · p. 3
• Additional analyses cover ablations with statistical testing, data-size scaling, frequency-band contributions, phoneme distribution shifts across splits, and phonetic-feature decoding.Source: §6 · p. 3

7.1. Ablation study and statistical significance

• Ablations compared variants differing by one change, using ten-seed means and standard deviations and paired Wilcoxon tests with p ≤ 0.01 as the significance threshold.This testing setup was used to distinguish component effects from random variation.Source: §7.1 · p. 3

7.1.1. Speech task ablation

• Extending speech windows from 0.5 to 2.5 s gave the largest speech ablation gain (+10.8% relative), and Conformer exceeded the SEANet baseline by +9.0%.Both effects were significant at p = 0.002; reducing stride to 60 also significantly helped (+2.8%), while removing augmentation had no significant effect.Source: §7.1.1 · p. 3, Table 2 · p. 4

7.2. Data-size ablation

• Speech Detection performance rose rapidly with more training data before beginning to saturate, while Phoneme Classification kept trending upward without an apparent plateau.The phoneme data-size analysis used examples without sample averaging; shaded bands show standard deviation across random seeds.Source: §7.2 · p. 4, Figure 4 · p. 4, Figure 5 · p. 5

7.3. Frequency band contribution to the tasks

• Speech decoding was above chance in Delta, Theta, Alpha, and Beta, while phoneme decoding was above chance in Delta, Theta, and Alpha but not higher bands.The speech task’s best single bands, Alpha and Delta, retained about 87% of full-band baseline performance; the phoneme task’s band-limited models remained below its unfiltered baseline, with averaging potentially suppressing higher frequencies.Source: §7.3 · pp. 4–5, Figure 6 · p. 5

7.1.2. Phoneme task ablation

• Dynamic grouping improved phoneme-task scores by +13.3% relative versus fixed groups (W = 1.0, p = 0.004), the largest reported phoneme ablation gain.Inverse-square-root class weighting improved scores by +7.6% but was not significant (p = 0.084); the custom Conformer’s +0.5% lift over Conformer Small was also not significant (p = 0.844).Source: §7.1.2 · p. 4, Table 3 · p. 5

7.5. Phonetic-feature decoding as a linguistically grounded probe

• In the non-averaged phonetic-feature probe, Voicing approached 58% F1-macro, Plosive and Fricative were moderately decodable, and Affricate remained at chance.Ensembling, transfer learning, and larger effective batches did not resolve the under-represented Affricate category; the authors conclude binary feature decomposition alone is insufficient to overcome data scarcity.Source: Table 4 · p. 6, §7.5 · p. 6

7.4. Distribution shift across splits in the phoneme task

• The phoneme holdout RMS distribution was bimodal with mean 0.64 and standard deviation 0.22, unlike the tighter unimodal validation and test distributions.This split shift motivated instance-level normalization to remove window-by-window amplitude and scale differences and improve holdout robustness.Source: §7.4 · p. 5, Figure 7 · p. 6

• The models achieved top-10 leaderboard performance on both tasks and won the LibriBrain 2025 Phoneme Classification track.Source: §8 · p. 6
• Instance-level normalization substantially improved Phoneme Task holdout generalization by mitigating distribution shifts between splits.Source: §8 · p. 6
• Both tasks drew most discriminative power from Delta–Beta frequency components, while Gamma and HGA contributed little in isolation.Source: §8 · p. 6
• SpecAugment-based augmentation benefited speech detection but showed limited benefit for the more complex, imbalanced phoneme task.Source: §8 · p. 6
• Because LibriBrain is single-subject, cross-subject generalization remains an open challenge.Source: §8 · p. 6
• Speech decoding began to saturate with additional data, whereas phoneme classification continued improving, suggesting potential gains from larger MEG datasets.Source: §8 · p. 6
• Future directions include sequential CTC or seq2seq objectives for speech reconstruction and linguistically grounded feature representations for interpretability and imbalance.Source: §8 · p. 6