Projects

Discovering Web Music using Signal Analysis

Discovering Web Music using Signal Analysis asks users to supply a source MP3, WMA or AAC from the web and plays a list of 50 tracks which are acoustically similar.

This demo provides some insight into the problems of working with genre. Many tracks within a "genre" often sound quite different and sometimes tracks that are in different genres sound very simliar. By ignoring the concept entirely, we can focus the discovery experience on finding new music which creates a similar "vibe" or "mood". As it happens, tracks which are acoustically similar appear to also create a similar vibe or mood.

Analyzing and then organizing web media tracks using signal analysis has many advantages. It allows us to discover tracks with little, incorrect or no metadata that we would otherwise never know about and unlike collaborative filtering it is able to recommend new tracks in the system immediately.

It's also completely egalitarian, recommending music simply on its similarity to the source track provided, rather than how popular, how well rated, or how well promoted the music is.

For this reason, this technique is perhaps most effectively used when applied to the ever increasing supply of user generated content. People listening to new music they have not heard, but which is similar to music that they do know and like.

The on demand nature of browsing by "stuff that sounds like this" also makes music discovery by signal analysis a highly personalized and engaging end user experience.

Discover New Music

Measuring Playlist Diversity

I worked with Malcolm Slaney from Yahoo! Research on measuring playlist diversity and trying to understand and characterize people's musical interests.

We looked at the 500 most popular playlists at Webjay and analyzed over 25000 audio tracks using MARSYAS, the open source software framework for audio processing.

Our work was accepted for publication and we presented our results at ACM Multimedia 2006.

Yahoo! Playlists

I am currently working with Lucas Gonze, founder of Webjay.org on the new Yahoo! Playlists site. We're taking the best parts of Webjay and mixing it together with the latest Yahoo! Music technology. The new site should go live early next year and will be very cool.

Yahoo! DJ

Yahoo! DJ is an award winning (the only Yahoo! Music Engine plugin contest winner to be publicly released) plugin I built which transforms the Yahoo! Music Jukebox into an exciting live music performance tool. It can be used to mix music at the club, the house party, your living room or even your workplace. It's a great way to really get to know your music - and have loads of fun along the way!

Get the Yahoo! DJ plugin

Web Media Browser

The Web Media Browser provides a web media discovery and browsing experience. I built it so I could surf the web for music from within YMJ. You can look for tracks in the Yahoo! Audio Search database, access the Webjay playlist community, browse the Magnatune music catalog or listen to what's appearing on various lists of the most recently blogged audio tracks. It even made it into Lucas' media browser cladogram

Get the Web Media Browser plugin

Sound of Markets

Music composition driven by realtime financial market information

Sound of Markets is a computer program I wrote in Java which composes and plays music based upon a streaming feed of numbers downloaded over the Internet, typically representing the bid or ask price for a financial instrument.

The program breaks down the datastream into differents sets of information representing the major trends for different time periods. For example, the bassline would be generated using a moving average over the past 22 days of data. The keyboard line is generated using the previous 5 days of data and the lead line is generated using the previous 5 hours of data. Depending upon whether the trend is positive or negative, different scales will be used for each instrument. The tempo of the music is determined based upon the current volume of activity, relative to the normal volume. On top of it all, up and down ticks trigger samples - with a series of up or down ticks triggering different samples.