REAL-TSE Challenge Dataset
收藏资源简介:
REAL-TSE数据集是由IEEE SLT 2026挑战赛推出的首个面向真实对话场景的目标说话人提取基准数据集,由南京大学、香港中文大学(深圳)等多机构联合构建。该数据集包含总计6,991个混合语音-注册语音对,涵盖11.3小时的真实汉语和英语对话音频,数据来源于AISHELL-4、AliMeeting、AMI等真实会议录音及新采集的多环境会话。数据集通过同步录制高质量麦克风、手机及耳机设备构建,保留了自然重叠、混响、噪声及信道失配等真实声学特性。该数据集旨在评估目标说话人提取技术在低延迟流式处理和全上下文批处理场景下的性能,推动语音分离技术在实际对话系统、助听设备和会议转录等领域的应用突破。
The REAL-TSE Dataset is the first benchmark dataset for target speaker extraction in real conversational scenarios, launched by the IEEE SLT 2026 Challenge and co-constructed by multiple institutions including Nanjing University, The Chinese University of Hong Kong, Shenzhen, etc. It contains a total of 6,991 mixed speech-enrollment speech pairs, covering 11.3 hours of real conversational audio in both Chinese and English, with data sourced from real meeting recordings such as AISHELL-4, AliMeeting, AMI, and newly collected multi-environment conversations. The dataset is constructed via synchronous recording using high-quality microphones, mobile phones and headphone devices, and retains realistic acoustic characteristics including natural overlapping speech, reverberation, noise and channel mismatch. It is designed to evaluate the performance of target speaker extraction technologies in low-latency streaming processing and full-context batch processing scenarios, and promote application breakthroughs of speech separation technologies in fields such as real conversational systems, hearing aids and meeting transcription.




