Fine pitch discrimination using Minim

Core Library Questions

oolong

Fine pitch discrimination using Minim

in Core Library Questions • 2 years ago

I've written a sketch that uses a FFT to find the peak frequency from a microphone input. It looks nice, but I'm having trouble finely distinguishing pitches in the frequency range where music and speech are usually packed. I'm aware of the rudiments of digital signal processing, but it's not obvious to me what settings would be best for this.

I currently have timeSize=2048, sampleRate=11025, which is reasonably responsive, but I don't think you'd want to tune your guitar by it. Am I missing a trick here? Is there some clever way of getting much higher resolution in ranges you're particularly interested in, or something?

Very grateful for any help.

Replies(3)

brokenbeep

Re: Fine pitch discrimination using Minim

2 years ago

Hi, not used Minim but I use FFT quite a lot. Firstly your using a pretty reduced sample rate, CD's generally use 44,100 so your sampling at a quarter of the resolution of a CD. If you increase the timeSize you increase the FFT resolution which will provide you with more frequency 'bins' and so a higher resolution, but to be fair 2048 is a pretty large window! With large windows like this you are analysing a broad range of frequencies but to the detriment of temporal change. In other words shorter windows are good for analysing sounds that change often, where capturing such change is important; and big windows are for more detailed analysis of longer sounds.

There is no way to isolate a single frequency range using FFT unfortunately, though you might want to look at wavelet transform...

oolong

Re: Fine pitch discrimination using Minim

2 years ago

Hi, thanks for that - and sorry I didn't respond sooner, for some reason I only just saw your reply!

I'm aware that I'm sampling at a quarter of the resolution of a CD - that's because as I understand it this lets me use the same window size to get four times the density of frequencies in this range (the A above middle C is only 440Hz, and I'm only interested in a maximum of two or three octaves either side - down to 110Hz, perhaps, up to 1760Hz). But I'm still a little hazy about this stuff...

jcobreros

Re: Fine pitch discrimination using Minim

2 years ago

Hi, I'm working on something similar and I'm having a related problem.

My app plots an FFT where the levels are in dB and the frequency scale is Logarithmic. This way I can plot melodies and bass lines as they would look on a piano. So far it's working fine for the higher pitched sounds (because higher notes cover more frequency bands, so it's more precise), but of course, any notes below 220-110hz shares frequency band with 5 other neighbours, so the bass lines are very poorly defined.

I read your post and you make a really good point about lowering the sampling rate to get more resolution on the lower end of the spectrum. I've tried lowering it to 22050 and i can see a change in resolution, similar to going from a 1024 to 2048 buffer size. Sadly it also introduces the same lag that going to 2048 would. I don't really understand why this happens. Have you had the same problem?

I'm using a very nice function called logAvg which divides de spectrum in octaves, so you can say logAvg(440, 12) and it will give you a plot where each band is a semitone starting at A 440hz. I have also tried using getBand and get the same result.

Apart from the 44100 vs 22050 problem, i've been thinking: each band is centered around a frequency, and has a bandwidth. Do the band frequencies overlap? Do we have some frequencies that are shared by two bands, and what's the weight curve? If so, maybe we can do some cool integration to get the values for each frequency more precisely.

Top Reply