EmphAssess
收藏资源简介:
EmphAssess数据集是由Meta AI Research和法国多家研究机构共同创建的,专注于评估语音到语音模型在处理和再现语音强调方面的能力。该数据集包含3652个合成语音样本,每个样本至少包含一个强调词,并附有详细的元数据,如转录文本和强调词的位置索引。数据集的创建过程涉及从内部文本到语音目的的手写转录本中选择转录本,并确保每个句子至少有两个不同版本的强调。EmphAssess数据集主要用于英语和西班牙语的语音到语音模型评估,旨在解决模型在跨语言和跨说话人转换中准确捕捉和再现强调的问题。
The EmphAssess dataset, co-created by Meta AI Research and multiple French research institutions, focuses on evaluating the performance of speech-to-speech models in processing and reproducing speech emphasis. This dataset comprises 3,652 synthesized speech samples, each containing at least one emphasized word, and is accompanied by comprehensive metadata including transcribed text and the positional indices of the emphasized words. The dataset's development process entails selecting transcripts from handwritten materials intended for internal text-to-speech initiatives, and ensuring that each sentence has at least two distinct versions with differing emphasis placements. The EmphAssess dataset is primarily employed for evaluating speech-to-speech models in English and Spanish, with the objective of addressing the challenge of models accurately capturing and reproducing speech emphasis during cross-lingual and cross-speaker conversion.




