Fredrik Cumlin

Gradient-based learning is now possible for variational sequential Monte Carlo without iterative solvers, thanks to a differentiable relaxation of systematic resampling.

Contrastive learning can significantly enhance speech quality assessment models without the computational burden of multi-stage training.

Overlooking high-frequency information in multi-rate speech can significantly hinder speech quality assessment, but SA-SSL-MOS shows how to leverage it for improved generalization.