QualiSpeech
收藏资源简介:
QualiSpeech是一个全面的低层次语音质量评估数据集,包含11个关键方面的详细自然语言评论,旨在通过丰富的注释来桥接自然语言反馈与语音质量评估之间的差距。该数据集由清华大学电子工程系创建,涵盖了人工合成语音和真实世界场景,提供了7个维度的数值评分和4个方面的具体描述。数据集通过综合注释过程生成,包括听众对语音样本的低层次特征进行评分和描述,以及利用GPT生成的自然语言描述。QualiSpeech旨在推动开发能够有效区分合成语音和真实语音的通用语音质量评估模型。
QualiSpeech is a comprehensive low-level speech quality assessment dataset containing detailed natural language reviews across 11 key aspects, which aims to bridge the gap between natural language feedback and speech quality assessment via rich annotations. Developed by the Department of Electronic Engineering, Tsinghua University, this dataset covers both synthetic speech and real-world scenarios, providing numerical scores across 7 dimensions and specific descriptions for 4 aspects. It is constructed through a comprehensive annotation pipeline, which includes listeners scoring and describing the low-level features of speech samples, as well as natural language descriptions generated via GPT. QualiSpeech aims to facilitate the development of universal speech quality assessment models that can effectively differentiate between synthetic and real-world speech.

- 1QualiSpeech: A Speech Quality Assessment Dataset with Natural Language Reasoning and Descriptions清华大学, 北京大学, Academia Sinica, 国立信息学研究所 · 2025年



