Musdb-XL-train

Here, we present the musdb-XL-train dataset for training De-Limiter networks. %%% Important Notes (2024-06-21) %%% We recently discovered some errors in the musdb-XL-train dataset. Specifically, about 7% of the training data (ozone_seg_0.wav ~ ozone_seg_20000.wav) had slight phase shift problems. If you are already using the musdb-XL-train dataset, please download the updated version. Sorry for the inconvenience. %%%%%%%%%%%%%%%%%%%%%% The musdb-XL-train dataset consists of a limiter-applied 300,000 segments of 4-sec audio segments and the 100 original songs. For each segment, we randomly chose arbitrary segment in 4 stems (vocals, bass, drums, other) of musdb-HQ training subset and randomly mixed them. Then, we applied a commercial limiter plug-in to each stem. Once you finish the download, you have to unzip it. The data is about 200~210GB so please be sure to make enough space. Due to the copyright issue, the dataset contains the sample-wise gain parameters (in .npy files), instead of a wave file itself, to make each wave file of musdb-XL-train data from the musdb18-HQ dataset. You should first prepare the musdb18-HQ dataset (https://zenodo.org/record/3338373). With the musdb18-HQ and this downloaded data (.npy and .csv), run the data processing code in our GitHub (https://github.com/jeonchangbin49/De-limiter, Please check the 'Musdb-XL-train' section). Then, you can get the actual wave files of musdb-XL-train data. After finishing the data processing step, you can remove the "np_ratio" folder that contains the sample-wise gain ratio parameters but you should keep your csv files because they will be used in our training process. Notice that our previous musdb-XL (https://zenodo.org/record/7041331) data is an evaluation dataset, and musdb-XL-train is a training dataset. --Dataset Construction For a commercial limiter plug-in, we used the iZotope Ozone 9 Maximizer, following our previous work, musdb-XL, which is a mastering-finished (in terms of a limiter, not an EQ) version of musdb-HQ test subset. The threshold parameters (related to the amount of a limiter operated) of the Ozone 9 Maximizer were chosen targeting the randomly selected loudness that sampled from the Gaussian distribution (mean -8, std 1). Parameters of the Gaussian distribution were selected following statistics of recent pop music (Refer the Table 1. of our previous paper, https://arxiv.org/abs/2208.14355). The character parameters (related to the attack and release parameters) of the limiter were randomly sampled from the gamma distribution (a=2, scale=1, in https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.gamma.html). The information on random mix parameters (gain and channel swap) is contained as csv files in our dataset.本研究面向De-Limiter网络的训练需求，发布了musdb-XL-train数据集。 %%% 重要说明（2024年6月21日） %%% 近期我们发现musdb-XL-train数据集存在部分问题：约7%的训练数据（ozone_seg_0.wav ~ ozone_seg_20000.wav）存在轻微相位偏移问题。若您已在使用该数据集，请下载更新后的版本。由此带来的不便，我们深表歉意。 %%%%%%%%%%%%%%%%%%%%%% musdb-XL-train数据集包含30万段经过限幅器处理的4秒音频片段，以及100首原始歌曲。对于每一段音频片段，我们从musdb-HQ训练子集的4组分轨（人声、贝斯、鼓组、其他音轨）中随机选取片段并进行随机混音，随后对每组分轨应用商用限幅器插件。 下载完成后，请对压缩包进行解压。该数据集大小约为200~210GB，请确保预留足够的存储空间。 鉴于版权问题，本数据集未直接提供音频波形文件，而是包含逐样本增益参数（存储于.npy文件中），用户需基于musdb18-HQ数据集生成musdb-XL-train的实际波形文件。请您先准备好musdb18-HQ数据集（下载地址：https://zenodo.org/record/3338373），结合该数据集与本次下载的.npy、.csv数据，运行我们GitHub仓库中的数据处理代码（https://github.com/jeonchangbin49/De-limiter，详见"Musdb-XL-train"章节），即可得到musdb-XL-train的实际音频波形文件。数据处理完成后，您可删除存储逐样本增益比参数的"np_ratio"文件夹，但需保留csv文件，因其将用于后续训练流程。 请注意，此前发布的musdb-XL数据集（下载地址：https://zenodo.org/record/7041331）为评估数据集，而musdb-XL-train则为训练数据集。 ——数据集构建 本次商用限幅器插件采用iZotope Ozone 9 Maximizer，延续了此前研究中针对musdb-HQ测试子集制作母带（仅限幅器环节，不包含均衡器处理）的musdb-XL数据集的制作方案。 Ozone 9 Maximizer的阈值参数（与限幅器的工作强度相关），其取值基于从高斯分布（均值为-8，标准差为1）中随机采样得到的响度值。该高斯分布的参数选取参考了当代流行音乐的响度统计特征（详见我们此前发表论文的表1，https://arxiv.org/abs/2208.14355）。 限幅器的特性参数（与起音、释音时间相关），从伽马分布（参数为a=2，scale=1，详见https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.gamma.html）中随机采样得到。 本次数据集的csv文件中，包含了随机混音参数（增益与声道交换参数）的相关信息。