Crowdsourcing enriches archives

Monday 21 Mar 16
The Danish Broadcasting Corporation (Danmarks Radio, DR) has somewhere in the region of a million hours of audio recordings, and desperately needs a system to put them in order. DTU Compute has algorithms that could help. 

Over the past ten years, a section of Denmark’s cultural heritage in the form of transmissions from DR has been digitalized. However, before this huge pool of material becomes genuinely useful, it has to be systematized and tagged in different ways so that searches can be run on it. One criterion is the identity of the person speaking, and work is being done on this aspect in the CoSound project, where DTU is one of the partners. Professor Jan Larsen from DTU Compute relates:

“We’ve built a system that can cut out the speech and separate it from music and jingles. The talking ‘bites’ are analysed for distinctive features, and a machine-learning algorithm can then be used to predict comparable voices.  The machine itself cannot identify the different speakers, but it can learn to predict who spoke a given ‘bite’ if it is given a few names to work with.”

“But who has the time to ‘teach’ the machine by listening to all the different ‘bites’ from start to finish? And how can we identify all the voices, not just the most famous ones? This is where we need assistance from a great many people, and the entire population of Denmark are experts in principle. So we came up with the idea of making the assignment interesting by ‘coating’ it with a game layer.”

No sooner said than done, and VOXVIP (vox = voice, VIP = Very Important People) has now been launched. You can choose between several different levels and compete against your self and other players to be the first to recognize voices. For each ‘bite’ of speech that typically lasts 10–15 seconds, you are given four possible answers plus a free text field where you can enter a name that has not been suggested. There is no ‘right and wrong’ answer sheet, but the system does award points for the answers.

“We have a great many algorithm variations that we have tested when people play the game. And you don’t even notice that you’re taking part in a research project designed to find the best algorithm for smart crowdsourcing, it’s simply a fun thing to do,” says Jan Larsen, who hopes that the Danish people will play along.