How AI Cover Songs Are Made: Vocal Isolation and Voice Conversion Explained
A music blogger outlines how AI cover pipelines work, breaking them into two chained machine learning problems: separating the vocal from a mixed track and converting the singer's voice. Vocal isolation is treated as an underdetermined problem, typically solved by neural networks that predict time-frequency masks, with tools like Deezer's Spleeter and Meta's Demucs cited as examples. The article notes common failure modes such as spectral overlap with instruments, reverb tails, blended backing vocals, and lossy source audio, which often require extra cleanup models. It also introduces voice conversion, which splits a vocal into content, pitch, and timbre components, and it points readers to open-source tools they can run themselves.
This is an AI-generated summary. ShortSingh links to the original source for the complete article.
Discussion (0)
Log in to join the discussion and vote.
Log in