Kaggle2025HackathonsResearch
Brain-to-Text '25
Decoding attempted speech from brain-implant neural signals into text. A pretrained RNN does the decoding, then an online n-gram model corrects what it gets wrong.
PyTorchRNNN-gram LMEnsembling
Status: In progress.
Objective
Decode a participant's attempted speech from brain-implant neural signals into text, scored on Word Error Rate (WER).
Approach
The competition ships a pretrained RNN baseline. Instead of retraining it from scratch, I treat its output as a first draft and put the effort into fixing that draft.
- Run the pretrained RNN to get raw decodes.
- Correct them with an online n-gram classifier acting as a language-model rescoring layer.
- Ensemble the RNN and the rescored outputs.
This is the one hackathon that sits closest to Tom. Turning a noisy signal into language is one of the sensory skills on Tom's list.
Where it's at
- ✅ Raw decodes from the pretrained RNN
- 🔄 Online n-gram rescoring layer
- 🔄 Ensemble weighting