Skip to content

Overview

Project Overview

Question: does an oscillatory inductive bias — specifically entrained oscillation, where phase and damping are input-dependent rather than fixed — improve phase reconstruction in monaural speech enhancement, compared to non-oscillatory (real-decay) and non-selective (LTI) SSM baselines?

Originated from the SSM-based architecture ideas in Ideas#Oscillatory SSMs for Phase-Aware SE.

Table of Contents

  • Project.excalidraw A visual overview of the project state and ideas.
  • Entrained Oscillatory State-Space Models for Phase-Aware Speech Enhancement Draft paper outline (abstract through discussion), plus a trailing Results section with the empirical comparison of LinOSS, S4D and Selective LRU as drop-in SSM cores in SEMamba (June 2026).
  • Results - July 26 July 2026 notes on S4D, LinOSS, Mamba and oscillatory Mamba parameterizations, with initialization results (e.g. real vs. rotating modes).
  • Theory (Selective LinOSS/Theory.md) Full theory of a selective LinOSS and the stability of SSMs in general.
  • S-LinOSS Proposal Proposal for a selective oscillatory SSM with a normal, scaled-rotation transition and a schedule-free stability guarantee.
  • SSM Domain Evaluation Potential project to run and assess different SSM cores across different domains of speech.

Status

LinOSS (as a drop-in LTI oscillatory core) has not shown a clear advantage over Mamba or a complex selective SSM (Selective LRU) for speech enhancement — see Results. The remaining open question is whether a genuinely selective oscillatory SSM (combining the harmonic-oscillator inductive bias with input-dependent phase/damping) does better; Theory works out how to make such a model stable.

As of September 2026:

  • Done: June 2026 comparison of LinOSS, S4D and Selective LRU as SSM cores in SEMamba; July 2026 parameterization and initialization experiments (Results - July 26); stability proofs for selective LinOSS (LinOSS Stability).
  • Next: run the factorial experiment of the entrained oscillatory S6 (rotation off/static/selective × damping static/selective, see Experimental Design), and validate the S-LinOSS surrogate against an LTI D-LinOSS baseline (caveat).
  • Not started: SSM Domain Evaluation.

Code, experiment tracking, machines and storage are documented in Infrastructure.