I’m trying to find an algorithm to detect when the user blows into the microphone (like Ninentdo DS or iPhone) on a WM6 device with C#.
I was thinking to use Waveform to read from the mic in real time into a small buffer, but I'm a bit lost on how to recognize the blowing sound, I guess it sounds like white noise. Anyone has an idea on how to do it?
Any help will be greatly appreciated.
If I were to tackle the problem, I'd blow into the mic and record that. Then run an FFT on the wave data to get a dominant frequency (or a few). I'd do that for several "blows" to see if they produce similar results. If they do (and I'm fairly sure they will) then I'd use an algorithm to look for that same frequency.
It's not a trivial task, but since a lot of the recording and math work is done, you could probably get it to work without too much pain.
This isn't exactly the same, but here is my answer to a question about note onset detection:
http://stackoverflow.com/questions/294468/note-onset-detection#294724
The answer describes an approach that relies on the signal strength's rising above a given threshold. This approach would work to detect a blowing noise, although it would also respond to any noise at all, so talking etc. would also trigger it.
You could use FFT, but I think this would be too slow to use in realtime, especially on a Windows Mobile device (it might be fast enough, though). However, a blowing noise is unlikely to have a dominant frequency, unless the blower is whistling a particular note. If FFT is fast enough, you might actually want to look for noises that don't have a dominant frequency (you'd be assuming that anything with a dominant frequency or frequencies was speech or an instrument of some sort) and that are heavily weighted towards the upper end of the audible range (above 10,000hz or so).
A blow in a mic will not necessarily show dominant frequencies in the high-range (10,00hz +). Blowing in a PC microphone will most likely cause signal staturation and distortion, which has a lot of low-range frequencies as well. The resulting signal will simply be a big burst of saturation.
I had to do a similar project years ago, and what I did was simply look for volume peaks well above normal speaking levels. Worked fine.
sthg's answer mentions distortion as a likely outcome of blowing into a mic. I just did a quick test, and my recording of blowing (on a cheap microphone) displayed a large number of samples at the minimum and maximum values (e.g. 32767 and -32768 on a 16-bits-per-sample WAV file), which looks like a flat-top has been applied to each of the peaks. A very simple way to detect the blowing sound might be to just count up all the samples that are at the min/max values, and if they represent more than 5% of the total (or some other threshold) assume that blowing is occuring.
I haven't looked at the spectrum of this, but I think you should see a band-limited noise spectrum with most energy below a few hundred Hz. This means you don't need a full audio bandwidth FFT, and 16K points is overkill. Even on a 256 point FFT over a small bandwidth you should be able to tell the noise from blowing in the microphone apart from speech.