I am trying to implement a simple VAD algorithm (a weighted ES VAD method from http://ceur-ws.org/Vol-2259/aics_17.pdf) and have been able to derive a mask that is either a 1 if speech conditions for that frame are met, and a 0 otherwise as shown below.
The top plot is the mask, the bottom is hand annotated (One small step for man audio). The problem I have not been able to solve is applying this mask to the audio. The frames are 25ms long repeated at intervals of 10ms (so 15ms overlaps) with F_sampling = 11025. I have tried applying the mask to the windows directly and then reconstructing, but the overlaps are causing a problem. Any ideas or pointers or criticism will be greatly appreciated.