Speaker Separation and Tracking System

Replicating human hearing in electronics under the constraints of using only two microphones (even with more than two speakers) and the user carrying the device at all times (i.e., mobile device weighing less than 100 g) is nontrivial. Our novel contribution in this area is a two-microphone system that incorporates both blind source separation and speaker tracking. This system handles more than two speakers and overlapping speech in a mobile environment. The system also supports the case in which a feedback loop from the speaker tracking step to the blind source separation can improve performance. In order to develop and optimize this system, we have established a novel benchmark that we herewith present. Using the introduced complexity metrics, we present the tradeoffs between system performance and computational load. Our results prove that in our case, source separation was significantly more dependent on frame duration than on sampling frequency.