Research map / Large language models / Language, safety and interpretability
HPSU: A Benchmark for Human-Level Perception in Real-World Spoken Speech Understanding
Chen Li, Peiji Yang, Yicheng Zhong, Jianxing Yu, Zhisheng Wang, Zihao Gou, Wenqing Chen, Jian Yin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| compare best model to humans | Gemini 2.5 Pro · HPSU average | accuracy: 87.3% (human) vs 62.6% | Hurt |
| use open-source Qwen2.5-Omni instead of proprietary Gemini 2.5 Pro | Qwen2.5-Omni · HPSU average | accuracy: 62.6 vs 60.0 | No clear effect |
| move from basic recognition (gender) to high-level reasoning (Scene, Mismatch Hard) | Speech LLMs · HPSU | accuracy | Hurt |
| use Whisper+GPT-4o cascade instead of end-to-end audio model on acoustic-detail tasks | Whisper+GPT-4o · HPSU EN Description | accuracy: 49.0 vs 51.6 (Gemini 2.5 Pro) | Hurt |
| use Whisper+GPT-4o cascade on semantic tasks | Whisper+GPT-4o · HPSU EN Intent | accuracy: 78.5 vs 79.4 (Gemini 2.5 Pro) | No clear effect |
| add misleading or hint-bearing prompts (induction) | all evaluated speech LLMs · HPSU | accuracy spread across prompt types | Hurt |
| use native omni-modal design or stronger reasoning | Gemini models · HPSU induction robustness | accuracy spread | Improved |
| evaluate in Chinese rather than English | human evaluators · HPSU | accuracy | Improved |
| switch language from English to Chinese for subtext | speech LLMs · HPSU Subtext | accuracy | Hurt |
Closest papers by meaning
- MMSU: A Massive Multi-task Spoken Language Understanding and Reasoning Benchmark (ICLR 2026 · similarity 0.92)
- Dynamic-SUPERB Phase-2: A Collaboratively Expanding Benchmark for Measuring the Capabilities of Spoken Language Models with 180 Tasks (ICLR 2025 · similarity 0.88)
- Listen, Think, and Understand (ICLR 2024 · similarity 0.88)
- SD-Eval: A Benchmark Dataset for Spoken Dialogue Understanding Beyond Words (NeurIPS 2024 · similarity 0.87)
- SVHalluc: Benchmarking Speech-Vision Hallucination in Audio-Visual Large Language Models (CVPR 2026 · similarity 0.87)
- Audio Large Language Models Can Be Descriptive Speech Quality Evaluators (ICLR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
