Source Separation AI, explained
Source Separation
What is Source Separation?
Source separation is a term used in the field of signal processing to describe the process of dividing an audio signal into different source audio signals. This idea applies to voice and machine noises in addition to music. To be able to edit the voices individually, you might want to separate the voices of two speakers in a podcast, for instance.
Why is Source Separation Such a Challenge?
Not everybody can play music. Even fewer people have a passion for data & AI while also being artists. It seems like non-musicians frequently believe that you can “take the voice and remove it from the audio” when you speak to them. This makes sense because there wouldn’t be thousands of karaoke renditions of well-known songs accessible at every pub, or instrumentals on the B-side of records. When you have access to the mix’s individual tracks, separating vocals from instrumentals is actually quite easy.
However, all we have in the real world are waveforms. The most accurate computer representation we have of an actual, physical audio event is a waveform. In order to convert digital audio back to real sound, such as through speakers, the waveform is also necessary. The combined waveform must be divided into two independent waveforms, each of which captures one of the sources accurately and exclusively, if you want to separate a piece of music into its two sources (vocals and instruments).
The image below shows three waveforms to emphasize this. The first one shows a guitar, while the second one shows vocals that are played over the guitar track. The guitar and vocals together, or the entire song, make up the third waveform.
Providing you with the vocals and instrumentation is a simple chore for me as the producer of this track because I can just email you the original recordings of both. However, as music listeners…
