Tag
RoleHear is a semi-supervised pipeline that enables a single cache-aware streaming transducer to emit speaker roles inline via lightweight in-vocabulary role tokens, using an offline teacher for pseudo-labeling and an LLM for silver role labels on unlabeled audio.
This paper systematically evaluates pseudo-labeling for adapting pretrained ASR models like Whisper and Qwen3-ASR to noisy police audio, introducing an LLM-as-a-judge filtering method and a cross-model paradigm to reduce word error rates.
The paper proposes label-free bias-only TTRL, which uses majority-vote pseudo-labels as rewards and optimizes only ~100K bias parameters with a frozen backbone, reaching 76.67% on MATH-500 with Qwen2.5-7B and transferring across vision-language and audio reasoning tasks.
StreamHear is a semi-supervised pipeline that adapts streaming speech recognition to domain-shifted audio by fine-tuning a teacher model on labeled data, generating pseudo-labels for unlabeled data, and fine-tuning a student model, with a realignment step for better word placement.
Selective Synergistic Learning (SSync) improves video object-centric learning by selectively distilling reliable cues via pseudo-labeling and transitive merging, avoiding error propagation from indiscriminate dense alignment.
SemiPrune is a label-efficient dataset pruning framework that uses semi-supervised learning to generate pseudo-labels from a small labeled subset, enabling existing supervised pruning methods to work with unlabeled data. It achieves state-of-the-art performance on domain-specific, image-corrupted, and long-tailed datasets.