Improper AMDF updated

one thing you can do with this ensemble is to drop the audio rate down to 11025 hz, as the FFT algo measures up to the sample rate, so you can get 1024 bins between 0-5000 hz, but at the cost of losing information on higher pitches. unfortunately these types of tradeoffs are everywhere with this type of work...

also, you guys may be interested in this
http://en.wikipedia.org/wiki/Cepstrum

although i've not figured it out quite yet.

 

I was just wondering about this the other day. If you could space the 1024 FFT bands logorithmically instead of linearly, this wouldn't be a problem at all.

There are ~134 semitones from 20Hz-20kHz which would give ~8 separate FFT bands per semitone -- more than enough for note identification and with some smoothing for glissandi.

The EzFFT Analyzer looks like it is in fact spacing its bands in some kind of curved way like this -- at least the graphical readout suggests so.
---

My impression is that any kind of useful polyphonic note detection will require sophisticated artificial intelligence pattern recognition of real-world overtone series and playing behavior. Just looking for the "loud bands" won't reliably be able to distinguish between overtones and different voices/instruments -- at different amplitudes in the mix -- with different fundamentals that are harmonious with each other.

I suspect even Melodyne doesn't do a good job with dense, complex source material, with notes starting at exact the same time. Try putting Nine Inch Nails "Head Like a Hole" in and see if you get a clean musical score out...

 

don't be fooled by the graphical readout - it's just graphing it logarithmically for looks, but it doesn't work that way to my understanding.

 

thank you so much for sharing your ensemble ! a lot of things becoming clearer to me now, while before checking out yours i was completely lost and don't even know where to start using the macros :|

that's interesting...we can lower just the internal clock within the FFT instead lowering audio rate i suppose... it even has positive impact on CPU usage !

anyway, i've tried quantizing the pitch output (while also lowering audio rate), and it seems very accurate on detecting the note number !

 

bert, i'm glad you found this useful. i was looking for a solid frequency detector to use with a traktor scratch record, which requires higher accuracy unfortunately.

i tried halving the core clock and i found that it made the pitch go up by an octave if the SRC was divided in half, which is pretty simple to work with..