Neural Keyword Spotting on LibriBrain

End-to-end walkthrough for Neural Keyword Spotting (KWS) on the LibriBrain MEG corpus: load data, frame the task, train a compact baseline, and evaluate with precision–recall metrics tailored to extreme class imbalance.

Overview

In the accompanying Jupyter Notebook tutorial for Neural Keyword Spotting (KWS) on the LibriBrain MEG dataset, you’ll run a compact temporal baseline, handle heavy class imbalance, and interpret precision–recall diagnostics (AUPRC, false alarms per hour at fixed recall).

Learning Goals

• Frame KWS from continuous MEG as a rare-event detection problem.
• Build and train a lightweight 1-D temporal model on consumer GPUs.
• Use AUPRC and FA/h at target recall to reflect user-facing trade-offs.
• Explore the first potentially useful application for non-invasive speech-BCIs!

Outline

1. Data access & labeling — load MEG windows and keyword labels; inspect prevalence.
2. Task formulation — windowing around events; train/val/test splits with safety checks.
3. Model — compact temporal conv + attention pooling; focal-loss training.
4. Evaluation — AUPRC, ROC (when defined), FA/h @ target recall; calibrated plots.

Notebook Access

• Anonymous GitHub (Filebrowser) here
• Direct Download here