Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Face-Voice Association with Inductive Bias for Maximum Class Separation

Marta Moscati, Oleksandr Kats, Mubashir Noman, Muhammad Zaigham Zaheer, Yufang Hou, Markus Schedl, Shah Nawaz

TL;DR

The paper defines face-voice association as a multimodal task and introduces maximum class separation as an inductive bias by applying a fixed matrix $P_{N_s-1}$ to the multimodal embedding $\mathbf{m}_i$, derived from frozen face and voice encoders. The model optimizes a combination of $\mathcal{L}_{CE}$ and $\mathcal{L}_{OC}$, with a weight $\alpha$, enforcing intra-speaker alignment and inter-speaker orthogonality. Empirically, it achieves state-of-the-art results on VoxCeleb seen-heard and unseen-unheard cross-modal verification and competitive performance on MAV-Celeb, demonstrating the first multimodal success of maximum class separation as an inductive bias. This work suggests a new paradigm for embedding discrimination in multimodal learning with potential broad applicability beyond classification.

Abstract

Face-voice association is widely studied in multimodal learning and is approached representing faces and voices with embeddings that are close for a same person and well separated from those of others. Previous work achieved this with loss functions. Recent advancements in classification have shown that the discriminative ability of embeddings can be strengthened by imposing maximum class separation as inductive bias. This technique has never been used in the domain of face-voice association, and this work aims at filling this gap. More specifically, we develop a method for face-voice association that imposes maximum class separation among multimodal representations of different speakers as an inductive bias. Through quantitative experiments we demonstrate the effectiveness of our approach, showing that it achieves SOTA performance on two task formulation of face-voice association. Furthermore, we carry out an ablation study to show that imposing inductive bias is most effective when combined with losses for inter-class orthogonality. To the best of our knowledge, this work is the first that applies and demonstrates the effectiveness of maximum class separation as an inductive bias in multimodal learning; it hence paves the way to establish a new paradigm.

Face-Voice Association with Inductive Bias for Maximum Class Separation

TL;DR

The paper defines face-voice association as a multimodal task and introduces maximum class separation as an inductive bias by applying a fixed matrix to the multimodal embedding , derived from frozen face and voice encoders. The model optimizes a combination of and , with a weight , enforcing intra-speaker alignment and inter-speaker orthogonality. Empirically, it achieves state-of-the-art results on VoxCeleb seen-heard and unseen-unheard cross-modal verification and competitive performance on MAV-Celeb, demonstrating the first multimodal success of maximum class separation as an inductive bias. This work suggests a new paradigm for embedding discrimination in multimodal learning with potential broad applicability beyond classification.

Abstract

Face-voice association is widely studied in multimodal learning and is approached representing faces and voices with embeddings that are close for a same person and well separated from those of others. Previous work achieved this with loss functions. Recent advancements in classification have shown that the discriminative ability of embeddings can be strengthened by imposing maximum class separation as inductive bias. This technique has never been used in the domain of face-voice association, and this work aims at filling this gap. More specifically, we develop a method for face-voice association that imposes maximum class separation among multimodal representations of different speakers as an inductive bias. Through quantitative experiments we demonstrate the effectiveness of our approach, showing that it achieves SOTA performance on two task formulation of face-voice association. Furthermore, we carry out an ablation study to show that imposing inductive bias is most effective when combined with losses for inter-class orthogonality. To the best of our knowledge, this work is the first that applies and demonstrates the effectiveness of maximum class separation as an inductive bias in multimodal learning; it hence paves the way to establish a new paradigm.
Paper Structure (6 sections, 3 equations, 2 figures, 3 tables)

This paper contains 6 sections, 3 equations, 2 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Method
  3. Experimental Setup
  4. Results
  5. Conclusion
  6. Acknowledgments

Figures (2)

  • Figure 1: Overall illustration of the proposed face-voice association method.
  • Figure 2: Performance of the models for cross-modal matching on VoxCeleb for varying gallery size $n_c\in\{2,4,6,8,10\}$. Left (a): comparison with SOTA methods. Right (b): ablation study on components of our method.