Fine-tuned Gaddy EMG-to-text checkpoint / recipe for the 33.95% WER result?
Hi Matteo, thank you for releasing TinyMyo and the silent_speech code.
I'm a student at MIT (Melbourne Institute of Technology in Australia) reproducing your EMG-to-text (recognition) results on the Gaddy
dataset as a baseline for some research. I've reproduced the from-scratch 4-layer setting
(~40.8% test WER, beam 1500) and I'm now fine-tuning from
pretraining/TinyMyo/TinyMyo.safetensors (8 layers) to reach the paper's ~33.95%.
Two questions, either would help a lot:
- Could you share the fine-tuned recognition checkpoint on Gaddy (the .pt that
start_training_from expects, i.e. the ~33.95% WER model)? I'd use it directly as a baseline. - If not, could you confirm the fine-tuning recipe: num_epochs / early-stopping patience,
LR schedule and peak LR, whether freeze_blocks was used, and the SizeAwareSampler token budget?
I've verified the backbone's transformer blocks load into the fork's EMGTransformer.
Thanks very much!
Erdenebayar Dovchindorj (MIT)
Hi Erdenebayar,
Thanks for reaching out, and great to hear that you reproduced the from-scratch baseline.
I have uploaded the fine-tuned checkpoint to the Hugging Face repository:
tinymyo_ft_emg2text_epoch_157.pt
You can run the evaluation of this checkpoint using the following command:
python recognition_model.py --model tinymyo --evaluate_saved tinymyo_ft_emg2text_epoch_157.pt
I also updated the configuration structure in the GitHub repository: https://github.com/MatteoFasulo/silent_speech
It would therefore be best to pull the latest main branch.
This is the 8-layer TinyMyo EMG-to-text checkpoint from epoch 157. It achieved 34.69% test WER. The reported paper result of 33.95% ± 0.97% is consistent with this checkpoint within run-to-run variation.
The fine-tuning recipe is the one available in the configuration file. There was no early stopping and checkpoints were saved after every epoch. Beam-search WER was evaluated every five epochs, while the training objective remained CTC loss.
Best regards,
Matteo Fasulo
thanks a lot