Perceptimatic

Perceptimatic数据集是由ENS/CNRS/EHESS/INRIA/PSL Research University的研究人员创建，包含法语和英语的语音刺激以及91名英语和93名法语听众的反应结果。数据集从自然阅读的语音语料库中提取，用于2017年Zero Resource Speech Challenge。该数据集旨在通过比较人类和模型的语音处理能力，评估无监督子词建模的效果。创建过程中，研究人员手动筛选和校正了5202个三元组，确保了数据的质量。Perceptimatic数据集主要应用于语音识别和人类语音感知的研究，特别是在无监督学习领域，以解决模型与人类感知差异的问题。

The Perceptimatic dataset was developed by researchers affiliated with ENS, CNRS, EHESS, INRIA, and PSL Research University. It includes speech stimuli in both French and English, alongside response data collected from 91 English-speaking listeners and 93 French-speaking listeners. Extracted from a naturally read speech corpus, this dataset was utilized in the 2017 Zero Resource Speech Challenge. Its core objective is to evaluate the efficacy of unsupervised subword modeling by comparing the speech processing abilities of humans and AI models. During the dataset construction process, researchers manually screened and corrected 5,202 triplets to ensure high data quality. The Perceptimatic dataset is primarily applied to research in speech recognition and human speech perception, particularly within the unsupervised learning domain, to address the discrepancy between model performance and human perceptual responses.