Atlas-grounded parcel-token architecture for decoding speech from intra-operative intracranial recordings across patients and sensor types. Extending Spalding 2025 (PCA+CCA, 8 patients, 9 phonemes, 0.31 balanced accuracy) with a scalable transformer architecture grounded in the Brainnetome parcellation on fsaverage.
Cogan Lab, Duke University · Ben Tang · with Greg Cogan (PI) and Zac Spalding
Parcels, not electrodes, are the shared representation across patients and sensors. Atlas-grounded parcel tokens should transfer cross-patient (within a sensor), cross-sensor (within an anatomy), and cross-lab (within a modality). No iEEG foundation model has tested cross-sensor representational transfer rigorously; that is the destination.
- Objectives — program hypothesis, stage roadmap, evaluation philosophy, advance gates
- Strategy — per-stage default architecture, frozen contract, live scoreboard (current:
stage_1.md) - Tactics — concrete task list, in-flight jobs, blockers
Full docs index: docs/README.md. Project conventions + code structure: CLAUDE.md.
Stage 1 (Phase 1) — single-sensor supervised correctness pass on intra-op uECoG, 7 LH patients (~7 min supervised). Architectural ablation wave in flight on DCC; Stage 2 (in-sensor scaling + SSL) blocked on 13 missing lexical FreeSurfer recons pending from Zac.
Python ≥ 3.11 managed via uv. pyproject.toml + uv.lock are authoritative.
uv sync # bootstrap .venv/
.venv/bin/python -m pytest tests/v14 -q # run test suite (338 tests)All training runs on Duke's DCC cluster, never local. See docs/references/dcc_setup.md.
src/speech_decoding/v14/— active per-phoneme Stage-1 implementationscripts/v14_core/— DCC sbatch wrappers + ablation log aggregatorscripts/ablation/— default DCC tooling (submit / status / logs / collect / peek / query / sync)docs/— triad + references + experiment logtests/v14/— test suite