Machine learning fuels Sun music recommendation technology

Open source project aims to overhaul music search


Software that listens to and analyzes music is driving a Sun project, which has a goal of creating an open source music recommendation system that surpasses the capabilities used today by iTunes and Amazon.

Software that listens to and analyzes music is driving a Sun open source project, which aims to build a music recommendation system that surpasses the systems used today by iTunes and Amazon

Automated recommendations today are typically based not on analysis of the music, but on who’s listening to which songs, said Paul Lamere, principal investigator of Sun’s Search Inside the Music. If one customer buys a Beatles album, he’ll be told what other music has been purchased by fellow Beatles fans.

Listeners’ preferences and buying habits vary widely, so these systems lead to some wacky recommendations. Lamere said one popular music service has told people who purchased the Britney Spears hit “Baby One More Time” that they might also enjoy an audio recording of the report on prewar intelligence by the U.S. Senate intelligence committee.

“It’s kind of a funny recommendation, and you have to think there’s something broken here,” Lamere said Tuesday at the Sun Labs Innovation Update in Burlington, Mass., where he demonstrated the Sun music search technology. Sun officials say they plan to release the software as open source, perhaps within six months.

Sun has built “one of the best music similarity algorithms” that’s based on the actual sound, with machine learning that analyzes features such as frequency and beats per minute to map out the rhythm structure, and determine the genre and which instruments are playing, Lamere said. Sun has taken advantage of prior research into speech recognition technology to tease out the features that correspond with the timbre of music and can be measured with computers, he said.

This technology could level the playing field between popular artists and newcomers who are trying to get attention in the increasingly crowded World Wide Web. It’s hard to find new and relevant music, Lamere said, because there are millions of tracks online and that number will expand into the billions, with the Internet acting as a repository for “the entire history of all recorded music.”

“Recommendation technology is key,” Lamere said. “The Web is going to be filled with billions of tracks and there’s going to be millions of tracks arriving every week. The question is, when you have a million songs in your in-box, how are you going to find something you really like?”

The project is reminiscent of Pandora, a free Internet radio service that recommends songs based on the results of the Music Genome Project, which analyzes sound based on hundreds of musical attributes.

Sound recognition technology is just one piece of Sun’s project, though. Sun’s other innovation is a tagging system that categorizes music based not on who’s purchased it but on its attributes, described with tags like “quirky,” “indie,” “rock,” “fast,” “frenzied,” “90s,” or “cute” and “fun.”

Sun is compiling these tags by searching reviews, lyrics, music blogs, social tagging sites and artist biographies, and incorporating the information into a prototype search engine Lamere demonstrated on Tuesday. Compiling the tags based on a comprehensive search of the Web prevents people from gaming the system by generating their own tags to enhance the popularity of certain tracks, he said.

Querying Sun’s prototype search engine for Led Zeppelin brings up a list of recommended artists – or “tagomendations” – such as Pink Floyd, Queen, the Rolling Stones and Jimi Hendrix. The user can then click a “why?” button to find out which tags overlap with the recommended artist and the one searched for. Hendrix is recommended for Led Zeppelin fans based on tags like “guitar gods,” “classic rock,” “guitar virtuoso” and “psychedelic.”

In addition to recommendations for other music, the search engine provides links to videos, pictures and upcoming concerts, if the artist searched for is alive and touring.

The tag system is better at analyzing music today than Sun’s machine learning algorithms.

“There’s so much you can’t get out of the audio,” Lamere said. “The fact that two songs played during your senior prom means a lot to you, yet it doesn’t" factor into sound recognition.

Still, computerized analysis of audio can fill in the gaps when artists are so new they simply aren’t being discussed on the Web, he said. In the demonstration, Lamere had Sun’s software “listen” to a Mozart piano sonata, and watched it spit out a list of recommended songs, mainly other classical music.

Future advancements to the computerized analysis will enable recognition of major and minor chords, bridges and choruses, and the rhythm patterns of reggae, pop and ska, Lamere said.

While Sun hasn’t perfected this type of high-level analysis yet, Lamere isn’t getting impatient. “I think I have the most fun job in the world,” he said.

Learn more about this topic

Virgin drops Windows Media-based music download service


Free music site attracts wrath of industry


ITunes Music Store pushes Sony aside on Yahoo Japan


Must read: 11 hidden tips and tweaks for Windows 10
View Comments
Join the discussion
Be the first to comment on this article. Our Commenting Policies