LibriCSS
收藏资源简介:
LibriCSS是由微软创建的一个多通道音频记录数据集,旨在评估连续语音分离算法。该数据集由LibriSpeech语料库的语句串联而成,模拟对话场景,并通过远场麦克风捕捉音频重放。LibriCSS包含10小时的音频记录,分为10个会话,每个会话约1小时。数据集设计考虑了不同的重叠比率和静音设置,以分析不同算法在各种重叠条件下的表现。此外,数据集还提供了地面实况分割,以便进行传统的语句级评估。LibriCSS的应用领域包括自动语音识别和说话人分割,旨在解决自然对话中语音信号连续性和部分重叠的问题。
LibriCSS is a multi-channel audio recording dataset developed by Microsoft for evaluating continuous speech separation algorithms. It is constructed by concatenating utterances from the LibriSpeech corpus to simulate conversational scenarios, with audio played back and captured using far-field microphones. LibriCSS comprises 10 hours of audio recordings, split into 10 sessions each lasting roughly one hour. The dataset is engineered with varying overlap ratios and silence configurations to analyze the performance of different algorithms across diverse overlapping conditions. Furthermore, the dataset provides ground-truth segmentation to support traditional utterance-level evaluation. Application scenarios of LibriCSS include automatic speech recognition and speaker diarization, targeting the challenges of speech signal continuity and partial overlap in natural conversational settings.




