Reproduce “Learning from Distribution: Rethinking One-Hot Supervision for ASR and QA in Speech Language Models” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yujie Guo, Hongjie Chen, Jiaming Zhou, Jian Kang, Jie Li, Yongxiang Li, Yong Qin
The authors use learning from distributions (LFD) to train speech models to match a text language model’s next-token predictions rather than learn only one transcript. In a shared-initialization full-parameter comparison, ASR-only cross-entropy achieved 5.52% word error rate versus 5.57% for LFD, while spoken-question accuracy was 1.89% versus 42.93%. This shows a QA difference with nearly unchanged recognition in this comparison.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.