SingSong: Generating musical
accompaniments from singing
|paper|
Chris Donahue*, Antoine Caillon*1, Adam Roberts*,
Ethan Manilow, Philippe Esling1, Andrea Agostinelli, Mauro Verzetti, Ian Simon, Olivier Pietquin,
Neil Zeghidour, Jesse Engel
Google Research, 1IRCAM, *Equal Contribution
Overview
We present SingSong, a system which generates
instrumental music to accompany input vocals,
potentially
offering musicians and non-musicians alike an intuitive new way to create music featuring their own voice.
To accomplish this, we build on recent developments in musical source separation and audio generation.
Specifically, we apply a state-of-the-art source separation algorithm to a large corpus of music audio to produce aligned pairs of vocals and instrumental sources.
Then, we adapt
AudioLM---a state-of-the-art approach for unconditional audio generation---to
be suitable for conditional ''audio-to-audio'' generation tasks, and train it on the source-separated (vocal, instrumental) pairs.
To improve our system's generalization from source-separated training data (where the vocals contain artifacts of the instrumental) to isolated vocals we might expect from users,
we explore a number of different featurizations of vocal inputs,
the best of which improves quantitative performance on isolated vocals by 53% relative to the default AudioLM featurization.
In a pairwise comparison with the same vocal inputs,
listeners expressed a significant preference for instrumentals generated by SingSong compared to those from a strong retrieval baseline.
