acophenotype: Pitt-trained models
Frozen models and reference data released with the PhD thesis Acoustic Phenotyping in Low-Resource Settings: A Multi-Representational Fusion Framework for Alzheimer's Detection
(Marek Sviderski, University of Sunderland, 2026). They are downloaded automatically by the
acophenotype Python packages; you do not
need to fetch them by hand.
pip install acophenotype
from acophenotype import AcousticProfile
profile = AcousticProfile.from_audio("recording.wav", task="binary")
Intended use and limitations
Research use only. Not a medical device and not for diagnosis or any clinical decision.
- Trained on the Pitt Corpus (DementiaBank): English, predominantly picture-description speech from a single study's recording setup.
- Performance has not been established for other languages, elicitation tasks, recording conditions or clinical populations.
- A single recording is a noisy observation; outputs should not be interpreted in isolation.
Contents
| Path | What it is |
|---|---|
fusion/stacking/ |
Late-fusion stacking (binary): logistic meta-model over per-stream P(AD) |
fusion/final_multiclass_sgl/ |
Sparse-group-lasso one-vs-rest model (multiclass) |
fusion/final_regression_fc2fs/ |
Correlation-filtered LightGBM regressor (regression) |
fusion/streams/ |
The per-stream base models used by the binary stacker |
fusion/biomarker_scaler.json |
Pitt z-score parameters applied to biomarkers before scoring |
reference/vlad_codebook.pkl |
16-cluster VLAD codebook for the Whisper embedding stream |
reference/*.json, umap_reference.npz |
Control-cohort statistics and UMAP layout for the report |
Each model has stream_only and with_demos (sex, age, education) variants.
Licence
The models and reference files are released under CC BY-NC-SA 4.0 (attribution, non-commercial, share-alike), consistent with the CC BY-NC-SA 3.0 terms governing the DementiaBank data they were trained on. They contain no audio or transcripts. The accompanying software is MIT-licensed.
Training data
Pitt Corpus, DementiaBank (TalkBank). Access to the corpus is governed by DementiaBank's terms; these files contain trained models and summary statistics, not audio or transcripts.
Environment
The models are serialized with pickle/joblib and are sensitive to library versions. They were saved with scikit-learn 1.7.2; the packages pin compatible versions. Staged with:
- python: 3.11
- scikit-learn: 1.7.2
- lightgbm: 4.7.0
- xgboost: 3.2.0
- numpy: 2.4.6
Pickle files execute code when loaded; load them only from this repository.
Citation
Please cite the thesis and the software: Sviderski, M. (2026). acophenotype. Zenodo. https://doi.org/10.5281/zenodo.22980874