Speech relation extraction (SpeechRE) aims to extract structured relational triples directly from speech. However, existing studies still face three major challenges: **the lack of Chinese SpeechRE benchmarks**, **insufficient modeling of speech-specific cues**, and **evaluation bias introduced by character-level exact matching** in Chinese scenarios. To address these challenges, we first construct **CSRTE**, a real-world Chinese SpeechRE benchmark built from human speech, providing unified data and evaluation settings. We then propose **ACRE**, a reasoning enhancement framework for SpeechRE, which integrates acoustic-cue and contextual-knowledge into a structured acoustic-semantic reasoning chain, enabling the model to go beyond literal semantic content **(what-to-say)** and capture speech-specific features **(how-to-say)** for relation reasoning. We further combine supervised fine-tuning with reinforcement learning and design composite rewards to improve the quality of the model’s reasoning process. In addition, we propose pronunciation-aware evaluation **(PAE)**, which considers pronunciation and tone consistency during entity matching to reduce the evaluation bias. Experiments on both Chinese and English SpeechRE benchmarks demonstrate the effectiveness of ACRE and the usefulness of PAE. Our source code and datasets are available at https://github.com/MWXGOD/CSRTE.

“中文语音实体关系三元组抽取任务(Chinese Speech Entity-Relation Triple Extraction Task, CSRTE)是第二十四届中国计算语言学大会中的一项技术评测,旨在从中文语音数据中自动识别并提取实体及其相互关系,构建结构化的语音关系三元组(头实体、关系、尾实体)。本任务的目标是提升中文语音关系三元组抽取的准确性与效率,增强模型在不同语境和复杂语音场景下的鲁棒性,实现从语音输入到文本三元组输出的全流程自动化处理。通过本次评测,有助于推动中文语音信息抽取技术的发展,促进语音与自然语言处理技术的深度融合,为智能应用提供更加丰富且精准的基础数据支持。此次评测共有257支队伍报名参赛,其中59支队伍提交了A榜成绩。成绩排名前15的队伍晋级A榜,并且表现突出的前朷支队伍提交了技术报告。”