According to The Verge, Google has developed an AI that can turn text suggestions into music that lasts for many minutes. In a research paper written by thirteen academics, Google recently unveiled its MusicLM AI model, similar to Open AI’s DALL-E picture generator.
The article contains a wealth of samples created using MusicLM, from meditation sounds and electronic music to death metal and rap, all of which are excellent five-minute recordings. The AI could also compose songs based on melodies and written instructions.
On one occasion, it created an operatic vocal to the hummed melody of “Bella Ciao.” Another example shows how MusicLM made a song from the word “gym” that featured illogical lyrical content and a vocal and melodic style characteristic of Arab pop.
One lengthy written suggestion suggests that the tool can produce a blend of reggaeton and electronic music “with a spacey, otherworldly sound,” which evokes the sensation of being “lost in space.”
Google has created a dataset of 5,500 music samples generated from a variety of text-based prompts, such as rich narrative captions, descriptions of famous paintings, instrument names, music genres, musician experience levels, places, and accordion solos https://t.co/KcGaq86npX
— Data Innovation (@DataInnovation) January 27, 2023
According to the study’s authors, “MusicLM exceeds earlier algorithms both in audio quality and conformity to the written description.” We also show that MusicLM can be trained on both text and a melody, allowing it to adjust the pitch and tempo of a hummed or whistled tune to match the tone of a captioned text passage.
- Google’s AirTag Rival: The Android Solution for Bluetooth Tracking
- Google Stadia’s Controller Upgrade and New Game Mark Shutdown
#Google detail #AI model MusicLM, which can generate high-fidelity music in any genre from text and was trained on a dataset of 280K hours of music 🎵
👉 https://t.co/prLegrzj9A@GoogleExpertUK pic.twitter.com/XTgxPnl31U
— Dr Efi Pylarinou (@efipm) January 28, 2023
Unfortunately, Google has “no plans to disclose models at this moment,” so you won’t be able to give the AI music generator a go just yet. The researchers warn of the dangers of “possible misuse of creative content” and the misunderstanding or distortion of cultural norms.
However, the study article announces the release of a public dataset including over 5,500 music-text combinations that Google claims will aid in the training and evaluation other music-based AIs.
Keep up with the latest news and information by following us on Twitter.