The Audio-to-Audio samples sound more like the model trying to construct a similar enough instrumental and vocoding the difference using the original sample rather than drums and strums being generated identically to match the intent. Still interesting though, and I certainly imagine we're not far off from improved versions.
No comments yet.