Posted on 21 November 2012
Part 3 in a series of videos recorded from ACM MIRUM 2012 in Nara, Japan.
Bryan Pardo presents an efficient method to equalize audio signals given user input in a customizable manner. For a given subjective acoustic property, e.g. "tinny", the user listens to several manipulated examples of the same audio signal and assesses the "tinniness" of each manipulated example. But without further optimization, this process is inefficient and time-consuming.
By using transfer and active learning, these subjective properties can be learned faster. First, each of the manipulated examples are different, and some of them assess tinniness better than others. Those examples which are best discriminate between "tinny" and "not tinny" should be assessed by the user before less discriminative examples. Second, if other users have already provided their opinions of tinniness to the system, then the system should be able to use that information constructively rather than begin the learning process from scratch. By using prior data, the system could learn the concept of "tinny" faster.