PhD research code for compact speech recognition via neural eigenfunction learning. Paper: arXiv:2604.06537 (see Citation).
| Task | Dataset | Result | Model size |
|---|---|---|---|
| Keyword spotting | TI-46 isolated digits | 99.5% accuracy | < 10 MB |
| Phoneme recognition | TIMIT | 22% Phone Error Rate | < 2M parameters |
The Functional Maximal Correlation Algorithm (FMCA) is a nonlinear feature extraction framework based on maximal correlation theory. Two neural networks are trained jointly with a log-determinant objective to recover the eigenfunctions of a normalized cross-density operator, maximizing statistical dependence between input and target variables without assuming any specific data distribution. The learned density-ratio eigenspace replaces correlation-based features and is well suited to prediction and classification on speech and other time-series signals.
Compared with conventional deep classifiers, FMCA features support markedly smaller downstream models: the keyword-spotting systems above stay under 10 MB after quantization while remaining competitive with much larger baselines (HMM and spiking neural networks on TI-46).
| Directory | Contents |
|---|---|
Classification/ |
Keyword spotting on TI-46 and Google Speech Commands (FMCA features + compact classifiers, quantized inference path) |
Randomization/ |
Synthetic-signal studies: Lorenz attractor and Mackey-Glass mixtures showing FMCA recovers the underlying pdf of two dynamical regimes even under full temporal randomization |
TCN/ |
Temporal Convolutional Network baselines for comparison |
KLMS_FMCA_Analysis/ |
Kernel adaptive-filtering analysis and comparisons |
If you use this code, please cite the accompanying paper (arXiv:2604.06537):
@article{sun2026time,
title={Time-Series Classification with Multivariate Statistical Dependence Features},
author={Sun, Yao and Hu, Bo and Principe, Jose},
journal={arXiv preprint arXiv:2604.06537},
year={2026}
}Released for academic and educational use. Please open an issue for other uses.