Functionally-grounded evaluation of dimensional interpretability in sparse speaker representations
Résumé
Dimensional interpretability is usually achieved by auto-encoding pre-trained embeddings into larger representations, optimizing a regularization term in addition to the reconstruction loss. One such regularization is sparsity, which has been described as a desirable property in interpretable representations. However, it is unclear how other regularizations, e.g. orthogonality, relate to sparsity, and how they compare w.r.t. interpretability. In this work, we adopt a functionally-grounded approach to measuring interpretability. To this end, we introduce a data-agnostic benchmark for dimensional interpretability applied to speaker representations (x-vectors), comprising a) unsupervised metrics, b) supervised disentanglement metrics, and c) a task-dependent metric as a guidance, i.e., typicality. We auto-encode x-vectors with different regularization strategies: SPINE (matrix-level sparsity), Top-K (vector-level sparsity) and DINE (orthogonality). With a wide assortment of such models, we show that sparsity and orthogonality are comparable strategies to train interpretable speaker representation models. They do not substantially degrade speaker verification performance and are in agreement with typicality, which was proposed as an interpretability measure for speaker verification. Finally, based on its agreement with typicality, we emphasize the relevance of MI-DCI as a supervised disentanglement metric for speaker characterization interpretability, which is in line with state-of-the-art conclusions on more controlled speech data.
| Origine | Fichiers produits par l'(les) auteur(s) |
|---|---|
| Licence |