MixAudio records the audio references used to generate individual musical parts. Generation is guided entirely by audio. In an October preprint, Neutune and KAIST researchers tested what those inputs contributed by replacing reference recordings or the surrounding mix.[1]
Changing a reference recording shifted measurements of the instrument’s sound colour towards that recording. Changing the surrounding mix moved harmony towards the replacement mix.[1]
An additional audit screened 994 generated melodic parts for resemblance to training recordings. Three listeners, all paper authors, assessed flagged cases. They identified one reproduced recording missing from the input record; 31 other matches arose from duplicate catalogue registrations.[1]
Only flagged candidates received listening checks, and bass and drums were excluded. The count is therefore a lower bound. The experiment distinguishes documenting a chosen source from checking what the output contains.[1]
Sources & further reading
- Tracing Inputs, Verifying Outputs: Validating Attribution in Music Generation Taejun Kim and colleagues / Neutune and KAIST · Accessed 29 Oct 2026