consistency-regularization

BranchShine-CR is a compact 25M-parameter model for multilingual IPA transcription that uses log-mel features, an E-Branchformer encoder, self-conditioned CTC, and consistency regularization, achieving a 22.3% relative reduction in IPA character error rate compared to prior models.

The paper identifies that reintroducing privileged context to a distilled student model degrades performance (context-induced degradation), and proposes a lightweight consistency regularizer that anchors no-context outputs to mitigate this issue, improving robustness across 12 configurations.

This paper proposes CAT, a cross-scale aligned transformer that enforces consistency between intermediate and final GAN outputs to resolve trajectory misalignment, achieving state-of-the-art FID of 1.56 on ImageNet-256.