Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Quantifying Dimensional Independence in Speech: An Information-Theoretic Framework for Disentangled Representation Learning

Bipasha Kashyap, Björn W. Schuller, Pubudu N. Pathirana

TL;DR

This work introduces an information-theoretic framework to quantify cross-dimension statistical dependence in handcrafted acoustic features by integrating bounded neural mutual information (MI) estimation with non-parametric validation and provides a principled framework for quantifying dimensional independence in speech.

Abstract

Speech signals encode emotional, linguistic, and pathological information within a shared acoustic channel; however, disentanglement is typically assessed indirectly through downstream task performance. We introduce an information-theoretic framework to quantify cross-dimension statistical dependence in handcrafted acoustic features by integrating bounded neural mutual information (MI) estimation with non-parametric validation. Across six corpora, cross-dimension MI remains low, with tight estimation bounds ($< 0.15$ nats), indicating weak statistical coupling in the data considered, whereas Source--Filter MI is substantially higher (0.47 nats). Attribution analysis, defined as the proportion of total MI attributable to source versus filter components, reveals source dominance for emotional dimensions (80\%) and filter dominance for linguistic and pathological dimensions (60\% and 58\%, respectively). These findings provide a principled framework for quantifying dimensional independence in speech.

Quantifying Dimensional Independence in Speech: An Information-Theoretic Framework for Disentangled Representation Learning

TL;DR

This work introduces an information-theoretic framework to quantify cross-dimension statistical dependence in handcrafted acoustic features by integrating bounded neural mutual information (MI) estimation with non-parametric validation and provides a principled framework for quantifying dimensional independence in speech.

Abstract

Speech signals encode emotional, linguistic, and pathological information within a shared acoustic channel; however, disentanglement is typically assessed indirectly through downstream task performance. We introduce an information-theoretic framework to quantify cross-dimension statistical dependence in handcrafted acoustic features by integrating bounded neural mutual information (MI) estimation with non-parametric validation. Across six corpora, cross-dimension MI remains low, with tight estimation bounds ( nats), indicating weak statistical coupling in the data considered, whereas Source--Filter MI is substantially higher (0.47 nats). Attribution analysis, defined as the proportion of total MI attributable to source versus filter components, reveals source dominance for emotional dimensions (80\%) and filter dominance for linguistic and pathological dimensions (60\% and 58\%, respectively). These findings provide a principled framework for quantifying dimensional independence in speech.
Paper Structure (25 sections, 14 equations, 4 figures, 1 table)

This paper contains 25 sections, 14 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Problem Formulation
  4. Feature Extraction
  5. Bounded Neural MI Estimation
  6. MINE with EMA Stabilisation
  7. CLUB with Variance Clamping
  8. KSG Non-Parametric Validation
  9. Combined Estimation with Adaptive Weighting
  10. Training Protocol
  11. Source-Filter Attribution
  12. Experimental Setup
  13. Datasets
  14. Implementation Details
  15. Results
  16. ...and 10 more sections

Figures (4)

  • Figure 1: Proposed framework for quantifying dimensional independence.
  • Figure 2: Cross-dimension MI (Final) heatmap across dataset combinations. Lower values (lighter) indicate greater independence between feature sets. Top marginal bars show per-pair means averaged across all combinations.
  • Figure 3: MI estimation convergence across dimensions: (a) Emotion–Linguistic, (b) Emotion–Pathology, (c) Linguistic–Pathology, and (d) Source–Filter. Per-ensemble MINE and CLUB trajectories are shown (offset for visual clarity), alongside the KSG baseline (dashed) and the final consensus estimate for a representative dataset combination (RAVDESS, L2-ARCTIC, UA-Speech). The mean estimator gap ($\Delta$) over the final 10 training epochs is reported, demonstrating progressive reduction and convergence over training.
  • Figure 4: Source–Filter attribution across semantic dimensions. Stacked bars show the mean proportion of MI carried by Source vs. Filter components (Eq. 3); individual dataset combinations are overlaid as jitter points. Error bars indicate 95% CI.