In a performance sensitive environment, I have an audio stream. I need to classify each frame as speech/non speech. For this purpose, only "clean" voice should be classified speech. Voice with substantial background noise (maybe music) should be classified as non speech.
What features/methods could you suggest?
I agree. I've previously researched this topic before and found that it is a very complex subject. Here are some basic algorithms: http://en.wikipedia.org/wiki/Speech_recognition#Algorithms.. I highly suggest you do lots of research. Here's some links for you to enjoy :) http://www.dsprelated.com/showmessage/83934/1.php & http://www.ee.columbia.edu/~dpwe/pubs/LeeE06-vad.pdf and many more on google.
What I would do: first try to find the fundamental frequency. A speaking voice does not have a fixed note in this sense, so you need to do it quite quickly-responding, a direct phase-locking method may be better than doing it with FFT. Then feed this frequency into a comb filter, to remove the fundamental and all its overtones. What remains is then, ideally, only pop and hiss noises, both either quite low or quite high-frequency, so bandpassing the midrange should – for a clear and single voice signal – leave only a very weak remaining signal. For music or other noises on the other hand, you have a wide mixture of frequencies throughout the midrange, so combfiltering will not weaken the RMS very much at all. So a high level after the comb/bandfiltering process will indicate that the source was not clean voice.
and it is not really reliable yet but does in principle work.
The combfiltered signal is 6 dB weaker than the only-bandpassed one.
Here, the combfiltered signal is actually louder (the filter is wrongly normalized).
This is not a simple problem, and this is more of a guess than a answer.
One distinction that comes to mind is that pure voice will have very little amplitude at higher frequencies, like above 3 kHz. Unfortunately, things like a hard S sound (hissing) does occur in voice and can have components up to around 8 kHz. Sometimes music and other background noise will have frequencies limited to 3 kHz too. So, frequency distinction will help, but isn't good enough on its own.
Lots of music will have a rythmic beat, especially when looking at just the base frequencies. That isn't true of all kinds of noise though.
As I said, this is not a simple problem and will probably require substantial experimentation.
My (long ago) Masters thesis involved doing this as a minor part of the overall task, largely with hardware. I can dig it out and see what I concluded :-). As I recall, the nature of energy content around 400 Hz is a major indicator. This was for use on telephone circuits. I also had some papers from British Telecom on the subject which I can provide references to and (just) possibly copies of the papers.
