Fluid Flow — Vocal Fold Transfer Learning

Fast ML regressors that predict vocal-fold acoustic outputs (fundamental frequency F0, sound pressure level SPL) from three motor inputs: cricothyroid activation a_CT, thyroarytenoid activation a_TA, and subglottal pressure PS. The point: train on a cheap source model (BCM) and transfer-learn to expensive targets (TBCM, Beam-Membrane FEM) to reduce expensive simulations.

Quick start

# Python deps
pip install scikit-learn torch tensorflow pandas numpy matplotlib joblib scipy

# --- Original female-BCM transfer (Brian) ---
python "VocalFoldRegression/BCM Model/RandomForest/MaleRF.py"
python "VocalFoldRegression/BCM Model/RandomForest/FemaleRFTransfer.py"

# --- BCM → BM transfer (Callum) ---
MPLBACKEND=Agg python Beam_Membrane/BM_TransferRF.py     # 6 RF methods
MPLBACKEND=Agg python Beam_Membrane/BM_TransferAE.py     # 3 autoencoder methods
MPLBACKEND=Agg python Beam_Membrane/BM_SmallData.py      # 10–500 sample sweep
MPLBACKEND=Agg python Beam_Membrane/BM_Summary.py        # cross-method comparison

# --- BCM → TBCM transfer (Callum) ---
MPLBACKEND=Agg python TBCM/TBCM_TransferRF.py
MPLBACKEND=Agg python TBCM/TBCM_Autoencoder.py
MPLBACKEND=Agg python TBCM/TBCM_Summary.py

Datasets:

• VocalFoldRegression/BCM Model/MaleBCM.csv — ~54k male BCM samples
• VocalFoldRegression/BCM Model/FemaleBCM.csv — female BCM (filtered ACFL > 30)
• Beam_Membrane/dataset_BM.csv — ~5,000 BM simulations (generated by Generate_BM_Dataset.m)
• TBCM/dataset_TBCM.csv — ~43k TBCM samples
• TBCM/dataset_TBCM_enriched.csv — TBCM + waveform features

CSVs are gitignored; rebuild from MATLAB or pull from your local data location.

Documentation

Operational (what we're doing right now — Brian + Callum + both Claudes)

Doc	Purpose
team/TODO.md	Master task list with owner + status. Single source of truth for "what is there to do"
team/BOARD.md	Kanban view (Backlog / In Progress / Review / Recently Done)
team/MEETING_NOTES.md	Append-only log of ~1pm syncs
team/README.md	Folder conventions, owner / status / priority values, cadence

Reference (what the project is)

Doc	Purpose
CLAUDE.md	Entry point, conventions, repo map (auto-loaded by Claude)
docs/ARCHITECTURE.md	System design, regressor matrix, transfer strategies (incl. autoencoder methods)
docs/MILESTONES.md	Dated history of what's shipped
docs/ROADMAP.md	Strategic research phases (multi-month)
docs/GLOSSARY.md	Domain terms, methods, file references
docs/DECISIONS.md	Append-only judgment log
PROJECT_GUIDE.md	Callum's hands-on guide for Beam_Membrane/ and TBCM/ (his standalone notes)

Branches

main and feature/fem are aligned. Work continues on feature/fem and is fast-forwarded to main after sync points.

Contributors

Equal collaborators; per-task ownership tracked in team/TODO.md. Original authorship of code areas:

• Brian Gladney — VocalFoldRegression/ (male/female BCM, RF/NN/PR baselines and transfer)
• Callum Camazzola — Beam_Membrane/, TBCM/ (BCM → BM and BCM → TBCM transfer; RF and autoencoder methods)