huggingworld/open-asr-leaderboard

该数据集名为ESB Test Sets: Parquet & Sorted，是一个用于自动语音识别（ASR）评估的测试集集合，基于open-asr-leaderboard/datasets-test-only数据集生成，主要特点是将数据按音频长度排序并转换为parquet格式以提高安全性和易用性。数据集包含8个子数据集：AMI（会议录音）、Common Voice（维基百科朗读）、Earnings22（财报电话会议）、GigaSpeech（有声书、播客和YouTube内容）、LibriSpeech（有声书）、SPGISpeech（财务会议）、TED-LIUM（TED演讲）和VoxPopuli（欧洲议会演讲），每个子数据集提供测试集，数据字段包括音频（采样率16kHz）、文本转录、数据集名称和唯一ID。数据集旨在简化ASR系统的训练和评估，支持通过Hugging Face Datasets库一键下载和准备，部分子数据集需要特定使用协议。此外，数据集还包含一个诊断数据集，用于跨领域评估。所有数据均经过错误校正处理，转录文本未提供测试集以支持公平评估。This dataset is named ESB Test Sets: Parquet & Sorted, a collection of test sets for automatic speech recognition (ASR) evaluation, derived from the open-asr-leaderboard/datasets-test-only dataset. It is characterized by sorting each split by audio length and converting the format to parquet for improved safety and usability. The dataset includes eight sub-datasets: AMI (meeting recordings), Common Voice (Wikipedia narration), Earnings22 (earnings calls), GigaSpeech (audiobooks, podcasts, and YouTube content), LibriSpeech (audiobooks), SPGISpeech (financial meetings), TED-LIUM (TED talks), and VoxPopuli (European Parliament speeches). Each sub-dataset provides a test split with features such as audio (sampling rate 16kHz), text transcription, dataset name, and unique ID. It is designed to streamline training and evaluation of ASR systems, allowing one-line download and preparation via the Hugging Face Datasets library, with some sub-datasets requiring specific usage agreements. Additionally, the dataset includes a diagnostic set for cross-domain evaluation. All data is error-corrected, and transcriptions are not provided for test splits to ensure fair assessment.