0
$\begingroup$

Actually i am in search of a Voice Activity Detection Algorithm which could distinguish between voice and non-voice

Roughly speaking it must not detect even a bullet sound,even a foot stepping and other non speech activity should only detect people conversation or any one shouting

in that search this question arises in my mind and i want know the effect of noise and distance between speaker and microphone on the speech features like pitch,frequency,cepstrum,zerocrossing rate,power spectral density,entropy etc

if some component wont get distorted i would like to extract that feature and do the activity decision on that

Can any one help me in extracting dominant parameter of speech which would differentiate it from other common sounds even in Lower SNR conditions <0dB

Note:my algorithm expects voice activity happens at a distance of at least 10m away from microphone and continuous generator hum as background noise

$\endgroup$
1
$\begingroup$

The intrinsic fundamental frequency (pitch) of a speech signal do not change when the speaker moves away from the microphone (unless it moves away quite fast in which case it's a case of Doppler effect!) - it would be strange if we heard people's voice transposed up or down as they move away or as ambient noise increases!

However, as the distance from the microphone increases, the signal to noise ratio will increase, and this might affect the reliability of a naive pitch detector, so you might indeed get different result when estimating the f0 of a speech signal and the same signal with more noise. It doesn't mean that the f0 is different, just that the noise makes the estimate unreliable.

I am not aware of any naive feature which allows noise-robust speech/non-speech discrimination with a simple threshold rule. I think supervised machine learning would be the way to go for your problem.

$\endgroup$
  • 1
    $\begingroup$ It is probably worth noting that although the fundamental frequency will not change as distance to the mic changes there will be a change in the spectral profile of the signal due to reflections and reverberations of the voice ending up at the mic. $\endgroup$ – PAK-9 Oct 10 '13 at 17:13
0
$\begingroup$

There is likely no one dominant parameter that is absolutely accurate. One possibility may be to extract a large number of parameters (as per your list, and more) from a very large corpus of example sounds that your detector might hear, and feed this data to a machine learning algorithm.

$\endgroup$
  • $\begingroup$ i am implementing this VAD on a DSP so i want it to be simple and things like machine learning will add up delay per frame in real time which become a serious problem $\endgroup$ – kakeh Sep 19 '13 at 10:33
  • $\begingroup$ I disagree. Machine learning techniques are expensive at the training stage, but this expensive phase will run on your computer, not on your embedded system. What will run on your embedded system is going to be as simple as a linear combination of (possibly transformed) features or maybe a tree of decision stumps... Something like hundreds CPU cycles per signal frame, far below the computational burden of extracting the features themselves in the first place. $\endgroup$ – pichenettes Oct 10 '13 at 20:31

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.