philgzl/dns5

DNS5挑战数据集是一个用于深度噪声抑制任务的数据集，是原始DNS5挑战数据的镜像版本。原始数据已从WAV格式转换为Opus格式，以减小文件大小并加速流式传输。数据集包含来自LibriVox、AudioSet、Freesound、OpenSLR26和OpenSLR28的数据，但排除了VCTK、VocalSet、CREMA-D、VoxCeleb2和DEMAND数据。技术规格包括：采样率48 kHz、单声道、Opus格式。数据集分为多个分割部分：英语语音（245小时，186743个文件）、法语语音（95小时，60454个文件）、德语语音（137小时，119175个文件）、意大利语语音（70小时，59525个文件）、俄语语音（25小时，16566个文件）、西班牙语语音（86小时，78603个文件）、噪声（177小时，63810个文件）、rir_slr26（17小时，60000个文件）和rir_slr28（0.1小时，248个文件）。许可证信息根据数据来源不同，包括公共领域、CC BY 4.0、CC0 1.0和Apache 2.0。数据来源为Microsoft的DNS-Challenge GitHub仓库，相关论文为ICASSP 2023 Deep Noise Suppression Challenge。DNS5 Challenge data is a dataset for deep noise suppression tasks, serving as a mirror of the original DNS5 Challenge data. The original files were converted from WAV to Opus format to reduce size and accelerate streaming. It includes data from LibriVox, AudioSet, Freesound, OpenSLR26, and OpenSLR28, while excluding VCTK, VocalSet, CREMA-D, VoxCeleb2, and DEMAND data. Technical specifications: sampling rate of 48 kHz, mono channel, Opus format. The dataset is split into multiple parts: speech_english (245 hours, 186743 files), speech_french (95 hours, 60454 files), speech_german (137 hours, 119175 files), speech_italian (70 hours, 59525 files), speech_russian (25 hours, 16566 files), speech_spanish (86 hours, 78603 files), noise (177 hours, 63810 files), rir_slr26 (17 hours, 60000 files), and rir_slr28 (0.1 hours, 248 files). Licensing varies by source, including public domain, CC BY 4.0, CC0 1.0, and Apache 2.0. The source is the Microsoft DNS-Challenge GitHub repository, with the associated paper being ICASSP 2023 Deep Noise Suppression Challenge.