Sure, it is relatively straightforward. Consider the various instruments of the orchestra. Each has a distinct mix of frequencies in the sound it is capable of producing. So we can distinguish a piano from a violin by examining the frequencies produced. Now consider that the human voice is just like an instrument with its own mix of frequencies, plus other characteristics like being a bit more staccato than a sustained instrument sound.

Mathematics in the form of a fast fourier analysis can decompose a sound into its component frequencies. We can also examine a stream using a windowing technique to divide up the stream into reasonable chunks.

All that remains is to use artificial intelligence in the form of a neural net to train a model, so that a window chunk of the stream fed into the NN model produces the output in the form of a likelihood of the input being a voice or not a voice. The more you can train the NN model, the more accurate will be the marks where the voices are.

--- (Edited on 2019-07-09 1:47 am [GMT-0400] by colbec) ---