The musdb18 is a dataset consisting of 150 music tracks, totalling to 10 hours of audio files. All files in the dataset are encoded in .mp4 and composed of 5 stereo streams: drums, bass, other instruments, vocals, and mixture, which is the sum of all the signals. The compressed stems as well as the uncompressed .wav files are available for use on Zenodo.

​​

To simplify the project so it can be done for a semester-long project, we only look at bass (percussion) and bass (harmonics). Bass sounds typically have lower frequencies and sustained tones. Drum sounds can vary across a broad range of frequencies, but certain components have characteristic frequencies (e.g., kick drums have lower frequencies and snares have higher frequencies).

 

We select the individual bass and drums files in a given music track, add them together, and separate the mixture back into bass and drums using different DSP techniques. The separated files can be compared with the original files quantitatively to determine which technique was the most effective.