PleasedPenguin/HEAR

HEAR 🎉 EMNLP 2026 main conference 🎉 📄 arXiv  ·  🌐 Project page  ·  💻 Code  ·  🤗 Dataset  ·  🧠 Model Hierarchical Evaluation of Attribution and Reasoning, a benchmark for speaker-attributed understanding of multi-party speech. Most speech benchmarks can be solved by transcribing the audio and reading the text. HEAR cannot. Every question asks something about who is speaking, not only what is said, and roughly half the benchmark comes… See the full description on the dataset page: https://huggingface.co/datasets/PleasedPenguin/HEAR.