You can deploy this model and start earning money today!

braindecode/
MetaNeuromotorHand

Views

No views yet

MetaNeuromotorHand

Generic neuromotor interface for handwriting from Meta (2025) [gni2025].

Architecture-only repository. Documents the braindecode.models.MetaNeuromotorHand class. No pretrained weights are distributed here. Instantiate the model and train it on your own data.

Quick start

pip install braindecode

python

from braindecode.models import MetaNeuromotorHand

model = MetaNeuromotorHand(
    n_chans=22,
    sfreq=250,
    input_window_seconds=4.0,
    n_outputs=4,
)

The signal-shape arguments above are illustrative defaults — adjust to match your recording.

Documentation

Full API reference: https://braindecode.org/stable/generated/braindecode.models.MetaNeuromotorHand.html
Interactive browser (live instantiation, parameter counts): https://huggingface.co/spaces/braindecode/model-explorer
Source on GitHub: https://github.com/braindecode/braindecode/blob/master/braindecode/models/meta_neuromotor.py#L34

Architecture

Parameters

Parameter	Type	Description
`n_outputs`	int	Vocabulary size for CTC. Defaults to `100` (handwriting charset).
`n_chans`	int	Number of EMG channels. Defaults to `16` (one armband).
`sfreq`	float	Sampling frequency in Hz. Defaults to `2000`.
`mpf_window_length`	int	MPF window length in samples.
`mpf_stride`	int	MPF frame stride in samples.
`mpf_n_fft`	int	STFT window / FFT size.
`mpf_fft_stride`	int	STFT hop size. Must divide `mpf_stride` and be `<= mpf_n_fft`.
`mpf_frequency_bins`	sequence of (float, float) or None	`(low, high)` Hz bands to average the cross-spectrum over. If `None`, all FFT frequency bins are used.
`mask_max_num_masks`	sequence of int	Max number of SpecAugment masks per dim (order matches `mask_dims`).
`mask_max_lengths`	sequence of int	Max mask length per dim (order matches `mask_dims`).
`mask_dims`	str	Axes to mask, among `"CFT"`. Defaults to `"TF"`.
`mask_value`	float	Filler value for masked regions.
`invariance_hidden_dims`	sequence of int	Hidden layer sizes of the per-rotation MLP. Output feature dim is `invariance_hidden_dims[-1]`.
`invariance_offsets`	sequence of int	Circular channel rotations to average over.
`num_adjacent_cov`	int	Number of adjacent off-diagonals of the cross-channel covariance matrix to keep.
`conformer_input_dim`	int	Conformer embedding dimension `D`.
`conformer_ffn_dim`	int	Feed-forward hidden dim inside each block.
`conformer_kernel_size`	int or sequence of int	Depthwise-conv kernel size per block.
`conformer_stride`	int or sequence of int	Depthwise-conv stride per block. As a scalar, applied only to the last block (entire encoder downsamples by `stride`); as a sequence of length `conformer_num_layers`, applied per block. Defaults to the paper's 15-layer schedule `(1, 1, 1, 1, 2) * 2 + (1,) * 5` (2x downsampling at blocks 5 and 10). When overriding `conformer_num_layers`, also pass a matching schedule or a scalar.
`conformer_num_heads`	int	Number of attention heads.
`conformer_attn_window_size`	int or sequence of int	Attention receptive field per block. Defaults to the paper's 15-layer schedule `(16,) * 10 + (8,) * 5`. When overriding `conformer_num_layers`, also pass a matching schedule or a scalar.
`conformer_num_layers`	int	Number of conformer blocks.
`drop_prob`	float	Dropout probability applied throughout the conformer (FFN, conv and attention blocks).
`time_reduction_stride`	int	Frame-stacking stride applied before the conformer. `1` disables it.
`log_softmax`	bool	If `True`, apply :func:`torch.nn.functional.log_softmax` to the emissions. Disabled by default (braindecode models return logits).
`activation`	type of nn.Module	Activation class used inside the conformer feed-forward and convolution blocks. Defaults to :class:`torch.nn.SiLU`.
`invariance_activation`	type of nn.Module	Activation class used inside the rotation-invariant MLP. Defaults to :class:`torch.nn.LeakyReLU`.

References

CTRL-labs at Reality Labs (Kaifosh, P., Reardon, T. R. et al.), 2025. A generic non-invasive neuromotor interface for human-computer interaction. Nature 645, 702-710. https://doi.org/10.1038/s41586-025-09255-w
Gulati, A. et al., 2020. Conformer: convolution-augmented transformer for speech recognition. Proc. Interspeech, 5036-5040.
Graves, A., Fernandez, S., Gomez, F., Schmidhuber, J., 2006. Connectionist temporal classification: labelling unsegmented sequence data with recurrent neural networks. Proc. ICML, 369-376.
Park, D. S. et al., 2019. SpecAugment: a simple data augmentation method for automatic speech recognition. Proc. Interspeech, 2613-2617.
Yu, J. et al., 2021. FastEmit: low-latency streaming ASR with sequence-level emission regularization. Proc. ICASSP.
Barachant, A., Barthelemy, Q., King, J.-R., Gramfort, A., Chevallier, S., Rodrigues, P. L. C., ... Aristimunha, B., 2026. pyRiemann (v0.10). Zenodo. https://doi.org/10.5281/zenodo.593816

Citation

Cite the original architecture paper (see References above) and braindecode:

bibtex

@article{aristimunha2025braindecode,
  title   = {Braindecode: a deep learning library for raw electrophysiological data},
  author  = {Aristimunha, Bruno and others},
  journal = {Zenodo},
  year    = {2025},
  doi     = {10.5281/zenodo.17699192},
}

License

BSD-3-Clause for the model code (matching braindecode). Pretraining-derived weights, if you fine-tune from a checkpoint, inherit the licence of that checkpoint and its training corpus.

braindecode/MetaNeuromotorHand