Within audio segmentation, I am specifically designing a system using a vector quantization approach on MFCCs to detect speaker changes in news pod-casts.
Taking an unsupervised approach I am aiming to design a "fast-forward" button between segments (or chapters).
My supervisor suggested an active learning approach, but I have been unable to find any previous work in this area of study.
I am thinking that active learning could be used to label possible candidates for speaker changes found using the aforementioned method. Thereby somehow designing a classification step for false alarm compensation.