Google’s logo is painted on the entrance to Google’s offices in London | Photo credit: Reuters
Researchers at Google have developed an AI tool, MusicLM, that can generate high-quality music from text descriptions.
(Subscribe to our technology newsletter Today’s Cache for insights on emerging themes at the intersection of technology, business and policy.)
MusicLM generates music at 24 kHz and stays consistent for several minutes. Both voice quality and adherence to text descriptions are better than previous systems, Google claims in a research paper Thursday.
The model can also convert whistling and humming melodies according to the style described in the text caption, Google added.
However, the tech company acknowledges there are some risks associated with the new model and the use cases it addresses, so it currently has no plans to release the model.
“We strongly emphasize that more work is needed to address these risks associated with music generation. At this time, we have no plans to release the model,” Google said in a study. .
The generated samples reflect the biases present in the training data, raising questions about the relevance of music production to cultures that are underrepresented in the training data, as well as concerns about cultural appropriation. points out Google.
According to the research paper, the company’s future efforts include generating lyrics, improving text conditioning and vocal quality, modeling high-level song structures such as introductions, verses and choruses, and modeling music at higher sample rates. may focus on.