PITCH DETECTION
kevin_thibedeau writes
There is a good way to make the
harmonic product spectrum more robust to the effects of noise. You
generate a synthetic spectrum starting from a histogram of zero crossing
intervals that has been "smeared" back into a continuous signal with
Gaussian peaks by applying a kernel density estimate. The result can
then be passed through the HPS to find the fundamental. The decorrelated
zero crossings caused by noise are much less problematic than working
with the HPS of the original signal.
Also, an overview here: http://obogason.com/fundamental-frequency-estimation-and-machine-learning/
Original link:https://news.ycombinator.com/item?id=10757879
PEAK DETECTION
John on comp.dsp writes (https://groups.google.com/forum/#!topic/comp.dsp/MqSJjoaq8Hg)
It is entirely inappropriate for some problems, but one peak-picker that I use sometimes is a three segment sliding window:
LLLLLLLLLLCCCCCRRRRRRRRRR
When the maximum of window C is in the middle of C, and that maximum
exceeds max(f(L),f(R)), then a peak at the middle of C is declared. The
function f can be mean(L) or max(L) or something else. The window
lengths are tuned to your requirements
Lito on comp.dsp writes
For a series with well behaved peaks the following
has worked for me. What it lacks in efficiency it
it makes up for with reliability.
function peaks(x,region,thresh)
value,index = max(x)
while value > thresh
add value, index to list
zero out region about index
value,index = max(x)
endwhile
return list
