Song search requires you to sing in tune

Australian computer scientists are developing technology that will enable fans to search for songs to download by singing into the computer.

Dr Sandra Uitdenbogerd, a researcher at Melbourne's RMIT University, says regular text searching is of little use if you don't know the exact title of a song, or if the song has common lyrics.

"If you know the words then you're far better off just typing some lyrics into Google and that will work brilliantly, but for certain types of music, that's not really an option," she said.

"In the classical music domain there are a lot of works where the words are all the same."

Uitdenbogerd, who holds a PhD in Computer Science and specialises in music information-retrieval technology, said she had been working on the project for 10 years.

"It's a logical next step that people become able to access music in a variety of ways and now that we have these large digital collections of music, singing is just one possible way of retrieving it," she said.

The technology could be ready for commercial use within three to four years and there was already commercial interest from "a local company that's still in the early stages at the moment", she said.

But ideally she would like to see the technology used by the most popular place for downloading music, which today is Apple's iTunes.

"If we're trying to index all the music in the world then maybe not in three or four years, but if it's just what's currently on the radio then maybe that's quite possible," she said.

Uitdenbogerd is working on the technology with several of her students and fellow staff at both RMIT and Monash universities. She was invited to speak about it at the recent Human Communication Sciences Network forum at Macquarie University.

All that's required to use the technology is a microphone, the relevant software and an internet connection, but the code behind it is extremely complicated.

Uitdenbogerd said working prototypes already allowed users to search for basic MIDI files, and the challenge now was to develop the software to the point where it could match voice signals with more advanced audio files like MP3s.

"What you sing into the computer gets stored as a wave form and in order to do matching against the music we have convert that into note information," she said.

"Then the next stage is to compare that to the note information that's been extracted from lots of audio files that are in a collection that you're searching."

A list of matches is then returned to the user, who can then choose which song to download.

But the tone-deaf may want to stick to basic text searches, as the less accurate your singing, the more difficult it is to find a match. Those with a sub-par singing voice may have to repeat their tune two or three times before the software has enough "information" to pinpoint the song.

"As it currently stands the transcription is very inaccurate, but the matching process compensates for that because it allows for errors," she said.