According to the document, the technology would work by first separating tracks into their component parts – vocals and instrumentals – then “adding the vocal component” of “at least one music track to at least a select segment of the base music track.” In other words, Spotify’s algorithm could automatically layer the singing from one song over the backing track of a completely different song, though users would still control the creative choices.
Source: Spotify holds a patent for tech that can generate song mashups