Advise how to find speech in long recordings (neural network?)