Midomi, a new company in Silicon Valley (Sunnyvale), has just launched an impressive musical search service.
If you hum, or sing a song, Midomi will likely recognize what you’re singing. So if you get a song in your head, and you don’t know what it is, go to Midomi. It will show you the most likely results, based on the signatures of songs it has in its database, and lets you go to a store to buy them.
That’s just the beginning. Midomi lets you see the people who have hummed the same songs, showing you those who sang most similar to you. It lets you chat with those people through a message box. It also lets you store your tune in its database for others to hear.
For example, see the screenshot below for Midomi’s results when I sang “Morning has Broken” into a microphone. At right, it shows that Cat Stevens was the original singer, and lets me buy the song at Amazon.com. At left you’ll see the profiles of the people who sounded most like me. At bottom, it gives other possible results. I can click on their profiles, listen and message them.
Midomi blows away Nayio Media, the San Mateo company that launched a humming search last month. If Nayio recognizes your tune, it cross-references with Napster’s library of songs. But the trouble is the recognition part. We tried humming “Morning has Broken,” but the song wasn’t represented on the first page of results. On Midomi, it was the first result, even when we hummed. Midomi gets better when you use words.
It was launched by two engineering PhDs from Stanford, Keyvan Mohajer (who studied music recognition), and Majid Emami, along with Stanford science grads James Hom and Michal Grabowski. The technology is self-built, called MMARS, or multi-modal adaptive recognition system. Midomi looks at pitch variation, like Nayio does, but it goes deeper, looking at tempo variation, phonetic content and location of pauses. If you hum a wrong note, it will often find ways to override it, by filtering it with other close signatures in its system. It isn’t perfect. While it recognized me humming “I Just Called to Say I Love you” immediately, it failed to do it on a later try when I hummed it quickly, without care. But once I added the words to my tune, it picked it up even when I sang quickly.
It gives you a studio, too, where you can pick out songs to sing — to help, it lets you play a 30-second track and see the lyrics — and lets you tag them, so that they are stored in the database. See a tour of the Midomi service here.
For copyright reasons, Midomi won’t let you upload your own songs, though that may change in the future. We did encounter a few bugs. Some sample tracks (it has licensed two million 30-second clips) didn’t play correctly (the audio didn’t come through). And one major caution: When we tried to download a software required before you can buy the songs, Midomi shut down the entire browser without warning. We still haven’t been able to download that software. We’ll update once Midomi fixes this.
The company’s goal is to become world’s most comprehensive database, Mohajer tells VentureBeat.
The Midomi service is their first product. They’ve formed a company called Melodis, and they want to move into other areas, such as speech recognition that will beat existing players like Nuance or Tellme. They’ve already implemented text recognition in Midomi. If type a search term “Ooops Britni Spirs,” it will recognize you are searching for the song by Britney Spears. Midomi also wants make the service easy to use for mobile devices.
The company has 15 employees in Sunnyvale, and another dozen developers in Eastern Europe and India. It has raised “multiple millions” from Global Catalyst Partners, and angel funding from Amidzad and former Googler, Aydin Senkut.
VB's research team is studying web-personalization... Chime in here, and we’ll share the results.