On the Distillation Loss Functions of Speech VAE for Unified Reconstruction, Understanding, and Generation

TL;DR

This paper systematically explores different distillation loss functions for speech VAE models, analyzing their impact on reconstruction, understanding, and generation, and proposes an adaptive joint-marginal alignment approach for improved performance.

Contribution

It introduces a comprehensive analysis of alignment strategies in speech VAE distillation and proposes an adaptive joint-marginal alignment method for better multi-task performance.

Findings

Joint-marginal alignment with adaptive weighting yields the best overall performance.

Different alignment approaches significantly impact speech VAE capabilities.

Adaptive weighting allows controllable balance among tasks.

Abstract

Continuous speech representations based on Variational Autoencoders (VAEs) have emerged as a promising alternative to traditional spectrogram or discrete token based features for speech generation and reconstruction. Recent research has tried to enrich the structural information in VAE latent representations by aligning with self-supervised learning (SSL) features, aiming for better generation performance. However, it remains unclear whether the widely-used alignment approach based on time-axis distillation is optimal when considering more tasks. To address this problem, this paper systematically explores different alignment approaches and analyzes their impact on the performances over three axes: reconstruction, understanding, and generation. We investigate various design choices in the distillation loss. Extensive experiments show that the joint-marginal alignment approach with adaptive weighting can achieve the best overall performance while allowing for a controllable balance.