Musdb18 Dataset

The musdb18 dataset was chosen as our starting point for our stem separation project. The data set consists of a total of 150 different songs split into a training subset consisting of 100 songs and a testing subset consisting of 50 songs. In total, there's approximately 10 hours worth of music from multiple genres. For each song in the dataset, we have access to a .wav file for the song in its entirety, which is a mixture of the constituent stems. Conveniently, we also have access to .wav files for the individual vocals, bass, drums, and "other" stems. The "other" stem is a mixture of any other instrumentation filling out the song. This dataset has proven to be extremely useful for our purposes as we have been able to start with the simple case of the separation of two stems and work towards the separation of three stems. Moreover, the individual stems can be treated as the ground truth for evaluating our source separation algorithms.

​

Each .wav file can be read into Matlab where it is represented by two vectors corresponding to the left and right channel of the .wav file respectively. In this fashion, each song is composed of five stereo streams: the mixture, the drums, the bass, the vocals, and the rest of the accompaniment. Each stereo stream is sampled at a frequency of 44.1 kHz. The mixture is separately encoded as AAC at 44.1 kHz and allows for a small difference between the sum of all individual sources and the entire song mixture. For this project, we are only considering the vocals, bass, and drums stems. By summing the individual stems together, we can create our own mixtures to feed into our source separation algorithms.

​

For the evaluation of our source separation algorithms, we focused on the vocals, bass, and drums mixture for six songs in the data set. Those songs are "Music Delta - Beatles", "Music Delta - Hendrix", "Music Delta - Rock", "BKS - Bulldozer", "Carlos Gonzalez - A Place For Us" and "Motor Tapes - Shore". We felt these samples were varied enough that they would be a good dataset to evaluate our performance. Attached below are the ground truth stems (vocals, bass, drums, and the three stem mixture) for each of these songs that we used for our testing. For the songs "BKS - Bulldozer", "Carlos Gonzalez - A Place For Us", and "Motor Tapes - Shore", we use a snippet of each song rather than the entire song to accelerate evaluation.