Using autocorrelation instead of calculating NSDF in MPM about pitch-detection HOT 5 CLOSED

Sure. Difference function and autocorrelation are quite similar.

In MPM, normalized square difference and autocorrelation functions are basically the same thing. I found in some exploration (https://github.com/sevagh/pitch-detection/tree/master/misc/mcleod#51---peak-picking-with-parabolic-interpolation---putting-it-all-together) that once you add the rest of the algorithm, autocorrelation and NSDF perform the same.

As a plus, I can use the same autocorrelation function in YIN (

from pitch-detection.

Thanks for explaining your approach in MPM.
Also, one more thing in YIN implementation remains unclear for me.

In the code above you are using:

ya->out_real[0] + ya->out_real[1] - 2 * ya->out_real[tau]; 

Why ya->out_real[1] is used here?

In my understanding this algorithm shoud look like:

T prefix_sum = 0;
for (int tau = 0; tau < ya->N / 2; tau++) 
{
    ya->yin_buffer[tau] = 2 * ya->out_real[0] - prefix_sum - 2 * ya->out_real[tau];
    prefix_sum += audio_buffer[tau];
}

So we add sum of all squares(ya->out_real[0]) and sum of all squares which start from tau(ya->out_real[0] - prefix_sum)

from pitch-detection.

I did that with trial and error.

You can see I used to have different code - similar to yours but missing the prefix_sum (maybe that's why it never worked): 521481b#diff-ec0aedadeb2a8475b194edb462daf8fcd235be1a347ae5d49feddbe7ede87559L112

from pitch-detection.

I would take a look at ya->out_real[1] (maybe print it). Perhaps by luck it's some approximation of the original algorithm.

from pitch-detection.

Thanks, for clarification. In the end, the best method is one which works :)

from pitch-detection.