LibriMix
收藏资源简介:
LibriMix是一个开源的语音分离数据集,由洛林大学和Inria联合创建。该数据集基于LibriSpeech和WHAM!的噪声样本,包含两到三名说话者的混合语音,旨在解决现有数据集在模型泛化能力上的不足。数据集包含约3000个混合样本,用于训练和测试语音分离模型。创建过程中,使用LUFS作为音量度量标准,确保混合语音的感知一致性。LibriMix的应用领域主要集中在提高语音分离模型在不同说话者和噪声环境下的泛化能力,特别是在实际对话场景中的应用。
LibriMix is an open-source speech separation dataset jointly created by the University of Lorraine and Inria. Built upon speech segments from LibriSpeech and noise samples from WHAM!, this dataset contains mixed speech from two to three speakers, aiming to address the shortcomings of existing datasets in terms of model generalization ability. It consists of approximately 3,000 mixed speech samples for training and testing speech separation models. During the dataset creation process, LUFS was adopted as the loudness metric to ensure perceptual consistency of the mixed speech. The main application scenarios of LibriMix focus on enhancing the generalization ability of speech separation models across different speakers and noise environments, particularly in real-world conversational scenarios.




