ericsyz/pa-pseudo-apa

该数据集是一个发音评估数据集，专为非母语者的开放式问题回答设计。它包含9,782个话语，来自1,003个不同的说话者，并提供了Gemini-2伪APA标签，包括准确性、流利度和韵律三个维度的评分。数据集被分割为训练集（6,109个话语）、验证集（2,673个话语）和测试集（1,000个话语），其中测试集与训练集和验证集在说话者上不重叠，以确保评估的独立性。音频数据以16位PCM WAV格式存储，未压缩大小约为8.6 GB。每个话语都附有整数评分，范围从1到5，分别对应准确性、流利度和韵律的评估。该数据集可用于发音质量分析和相关NLP研究。Pronunciation Assessment Dataset is a dataset for pronunciation assessment, featuring non-native open-ended question responses with Gemini-2 pseudo APA labels (Accuracy, Fluency, Prosody). It contains 9,782 utterances from 1,003 speakers, with splits into train (6,109 utterances), validation (2,673 utterances), and test (1,000 utterances) sets, where the test set is speaker-disjoint from both train and validation. The audio is in 16-bit PCM WAV format, approximately 8.6 GB uncompressed. Labels are integer scores ranging from 1 to 5 for Accuracy, Fluency, and Prosody per utterance.