CLMT: A Parameter-Efficient Foundation Model for Physiological Signal Synthesis

Researchers propose Compact Latent Manifold Translation (CLMT), a 0.09B parameter unified framework for cross-modal and cross-frequency synthesis of physiological signals like ECG and PPG. CLMT uses a two-stage discrete translation paradigm: a Universal Tokenizer with Hierarchical Residual Vector Quantization (RVQ) decouples heterogeneous signals into isolated discrete latent manifolds, preventing modality interference. A Context-Prompted Latent Transformer then enables high-fidelity generation across modalities and frequencies. The model addresses modality entanglement and high computational costs in existing foundation models, aiming for edge-device deployment. The paper is available on arXiv.