Isolating sections of human voice from noisy audio

Hello there all the DSP experts. As a novice with a problem, I have a humble request for guidance.

I am working with several 15-hour audio samples captured in an industrial environment of the following kind: Noisy background, constant whirring of machines, as if driving in a noisy vehicle, no major repetitive noise, some sudden sharp 'ping' (alarm signals from certain machines), and on top of that infrequent/sporadic human voice.

My main interest lies in hearing out the human communication. Currently, I am spending listening to the entire 15 hours!

I do not intend a perfect solution, just want to save as much time as possible, by trimming the portions where no human voice exists.

Is there any way I can achieve this through Signal Processing?

Thank you very much all.

[Amit]

Any starting points, if not perfect solutions will be highly appreciated. Thanks.

Answer 1

[Image Analyst]

I think you need to use Independent Components Analysis (ICA), also known as Blind Source Separation (BSS) to solve the "Cocktail party problem". I'm not an expert in that so all I can do is to refer you to these links:

http://research.ics.aalto.fi/ica/icademo/

http://research.ics.aalto.fi/ica/cocktail/cocktail_en.cgi

http://research.ics.aalto.fi/ica/fastica/

http://cnl.salk.edu/~tony/ica.html

[Amit]

Thanks Image Analyst. While I try this, is there any 'makeshift' approach? My 15 hour samples have typically 10-30 minutes of human voice. Can anything coarser may be done. Thanks again.