SingSong was created by researchers from the Google Research Institute. This artificial intelligence is described as “a new, intuitive way to create music” for more or less seasoned singers.
Jesse Engel and his colleagues used an algorithm to separate the instrumental and vocal tracks from 46,000 hours of music, then used an artificial intelligence model previously developed by Google Research to harmonize these different recordings. This experimentation phase allowed the AI to learn how to work with audio tracks that weren’t initially in its database, and create bespoke musical accompaniments, according to the review. New Scientist.
Are they at the level of those imagined by a human? To answer this question, the scientists played 800 short AI-enhanced recordings to music lovers. They wanted to know if they were able to differentiate between songs whose backing tracks had been generated by the machine, those where they had been composed by a real musician, or even added randomly.
Participants in the study, whose findings were published online, preferred the contribution of artificial intelligence in 66% of cases. They found that the musical accompaniments she had generated were more compatible with the vocalist’s or lead singer’s voice. “From a subjective point of view, SingSong produces instrumental tracks that have a clear harmonic and temporal correspondence to the input vocals”say the researchers.
A direct competitor to musicians?
They noticed, however, that SingSong had a harder time creating convincing backing tracks when the vocals were recorded by amateur singers. The reason ? The latter often struggle to stay in the right tempo and maintain the same pitch of voice. Jess Aslan, lecturer at the University of London, however, finds these initial results very convincing. “I wouldn’t say you could release the track as such. It would need to be tweaked, but for sure it seems to be working in terms of consistency”she told the New Scientist.
The creation of SingSong raises many questions about the role that artificial intelligence will play in the music industry in the future. Some fear that it will become a direct competitor to songwriters, producers, even singers. Indeed, Microsoft recently developed a tool called Vall-E, capable of reproducing any voice from a simple three-second sample. This innovation could potentially be used in the creation of new songs, even if it was not created for this specific purpose.
Be that as it may, many voices demand the establishment of firm regulations in the face of the success encountered by so-called generative artificial intelligences. China has taken the lead on this issue by announcing the entry into force on January 10, 2023, of a law requiring that all content generated by AI be clearly identified as such, according to TechCrunch. It’s a safe bet that other governments will follow its example in the coming months.