Music-Savvy AI Built at UMD Could Help You Discover Your Next Favorite Artist

A new AI model that understands music like a trained musician is one beat closer to reaching your devices.

Music Flamingo, developed by University of Maryland and NVIDIA researchers, is at the center of a recently announced partnership between NVIDIA and Universal Music Group (UMG) that aims to revolutionize how fans discover songs and how musicians create them. 

The partnership “unites the world’s leading technology company with the world’s leading music company in a shared mission to harness revolutionary AI technology to dramatically advance the interests of the creative community,” said Sir Lucian Grainge, UMG’s chairman and CEO, in a statement on the deal. 

Streaming services’ current music recommendation algorithms rely on clicks and user behavior rather than musical analysis; if you cue up a Taylor Swift song, they may recommend Sabrina Carpenter next because listeners who like the first tend to like the second. Instead, a Music Flamingo-driven algorithm could deliver personalized recommendations based on finer-grained listener preferences: stylistic, compositional, even emotional.

UMG and NVIDIA say the corporate partnership reflects the promise of Music Flamingo, which produces in-depth, coherent narratives about songs. Its musical interpretations are preferred by real musicians and outperform those produced by existing AI models in over 10 benchmarks assessing music understanding and reasoning. 

“When Music Flamingo came out and started beating everything else, the whole music industry got really interested,” said Ramani Duraiswami, a professor in UMD’s Department of Computer Science and a co-creator of Music Flamingo.

To illustrate the model’s performance, the researchers published several examples of Music Flamingo’s descriptions, including for Swift’s “The Fate of Ophelia”—the chart-topping lead single from the singer-songwriter’s latest album, “The Life of a Showgirl.” 

The model clocks the song’s tempo of 125 beats per minute and notes it’s in the key of F major. It details how the composition “blends bright, melodic synth-pop sensibilities with a polished, modern electronic production” and how Swift’s vocal “bright and breathy timbre” sits over “tasteful reverb and delay, creating a spacious, ethereal ambience.” Even the lyrics fall under the AI microscope: they “explore a narrative of rescue and transformation, using the metaphor of ‘Ophelia’ to depict a past of isolation and a newfound sense of purpose.”

Duraiswami credits the model’s impressive performance to its training, which includes over 4 million songs from multiple languages and genres. A detailed musical description accompanied each song, and the dataset included an additional 1.8 million question-answer pairs curated to instill musical reasoning. The process took 128 of NVIDIA’s most powerful GPUs a month to complete, and human musicians worked with Music Flamingo afterward to reinforce the model’s musical reasoning skills. 

Developing an AI model capable of holistic musical analysis is a watershed moment, said Sreyan Ghosh, a doctoral student in computer science at UMD and lead author of the Music Flamingo preprint posted on arXiv. He said much AI research on music has lagged because it’s been difficult to curate a large training database of full-length songs with accompanying in-depth descriptions. Plus, it’s challenging for AI to parse through the dense and layered information in music—from vocals to production to instrumentation. His team will present their work at the 2026 International Conference on Learning Representations, one of the world’s premier conferences in machine learning.

Beyond aiding in music discovery, Ghosh, whose research was supported by the NVIDIA Graduate Fellowship, sees the AI’s greatest potential in helping artists themselves. For instance, emerging artists could actively use the service for outreach to potential fans who are likely to engage deeply with their work. Music Flamingo could also help musicians safeguard their intellectual property. His team is currently brainstorming ways to detect the hallmarks of AI-generated music—and to pinpoint the copyrighted pieces from which AI songs lift. 

Music Flamingo could even help artists create music. UMG and NVIDIA, for instance, will establish an “artist incubator” to test AI-driven music creation tools that prioritize “originality and authenticity—serving as a direct antidote to generic, ‘AI slop’ outputs, and placing artists at the center of responsible AI innovation,” according to the January announcement.

Fully integrating AI into music creation and discovery will require building trust between two industries that have publicly clashed over copyright and artistic concerns. That’s top of mind for Ghosh as he prepares to join NVIDIA to work on Music Flamingo as a full-time employee after graduating this spring. 

“Anything that we build should be loved by the music creator family,” Ghosh said.

He’s cautious when predicting what the future might hold, but he envisions a world where people use AI to engage with music as casually as they use large language models today. That could mean that streaming services use AI to curate playlists catering to the listener’s taste and mood. If the perfect song doesn’t exist, they might even generate music—ethically, he noted—bespoke to what a listener wants. 

“I think the ChatGPT moment for music is yet to come,” he said. “That is where we feel that Music Flamingo can play a role—in doing everything to improve the music listening experience for the daily user.”