ASR-GLUE

ASR-GLUE是由腾讯AI实验室创建的一个多任务基准数据集，旨在评估自动语音识别（ASR）错误对自然语言理解（NLU）系统的影响。该数据集包含6种不同的NLU任务，涵盖了情感分析、语义文本相似性、自然语言推理等多个领域。数据集通过聘请6位不同背景的母语者，在3种不同级别的背景噪音下录制测试数据，共计产生了18种不同类型的音频记录。ASR-GLUE不仅用于评估模型的性能，还旨在推动ASR-鲁棒NLU的研究，解决实际语音交互系统中ASR错误导致的NLU性能下降问题。

ASR-GLUE is a multi-task benchmark dataset developed by Tencent AI Lab, which is designed to evaluate the impact of automatic speech recognition (ASR) errors on natural language understanding (NLU) systems. It encompasses six distinct NLU tasks spanning domains including sentiment analysis, semantic textual similarity, and natural language inference. Test data was collected by six native speakers with diverse backgrounds under three different levels of background noise, yielding a total of 18 distinct types of audio recordings. Beyond evaluating model performance, ASR-GLUE also aims to promote research on ASR-robust NLU, addressing the problem of degraded NLU performance caused by ASR errors in practical speech interaction systems.