Source separation is the process of taking an auditory mixture of sounds and isolating its individual sound components, referred to as a sources. The classical source separation problem is the cocktail party problem, in which several people are talking simultaneously in a cocktail party and a listener wants to identify one from the discussion. This technique has numerous applications, including music analysis, speech processing, and beamforming. Nowadays the source separation problem is commonly approached by machine learning methods, but in this project we also explore some traditional signal processing methods. We focus on music. The figure below illustrates this concept in the context of music.​

​​

​

​

​

​

​​

​

​

​​

 

Music is a complex mixture of instrumental sounds, such as bass, drums, piano, and vocals. Separating these sources poses unique challenges compared to other types of source separation. Musical instruments are highly correlated, as they all tend to change simultaneously and may harmonize. This correlation creates significant overlap in the time and frequency domains, complicating the separation process. Additionally, creative processing tools such as reverb and equalization are routinely used in music production but can make it difficult to separate the sources.

​

However, source separation remains an interesting problem with a wide range of practical applications. Because isolated sources are easier to process and manipulate than mixtures of sources, source separation is a valuable technique for music remixing, singer and lyric identification, and fundamental frequency estimation.