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

A team of researchers including Bipasha Kashyap, Björn W. Schuller, and Pubudu N. Pathirana has published a groundbreaking paper titled 'Quantifying Dimensional Independence in Speech: An Information-Theoretic Framework for Disentangled Representation Learning' on arXiv. The work addresses a fundamental challenge in speech processing: speech signals naturally encode multiple types of information—emotional state, linguistic content, and potential pathological indicators—all within a single acoustic channel. Traditionally, the success of disentangling these dimensions has been assessed indirectly through the performance of downstream tasks like emotion recognition or automatic speech recognition. This new research introduces a direct, principled framework to measure the statistical independence between these dimensions using information theory, specifically by estimating the mutual information between handcrafted acoustic features.

The core technical innovation is the integration of bounded neural mutual information estimation with non-parametric validation. Applied across six different speech corpora, the analysis yielded significant quantitative insights. It found that the statistical coupling (cross-dimension MI) between different information types (e.g., emotion and words) is surprisingly weak, with tight estimation bounds under 0.15 nats. In contrast, the mutual information between the source (vocal fold vibration) and filter (vocal tract shape) components of speech production was substantially higher at 0.47 nats. Crucially, the framework enables an 'attribution analysis,' quantifying whether the source or filter component carries more information for a given dimension. Results showed source dominance for emotional dimensions (80%) and filter dominance for linguistic (60%) and pathological dimensions (58%). This provides a concrete, measurable foundation for evaluating and improving disentangled representation learning models, moving the field beyond proxy performance metrics.