How does Kick Analysis work? - Make Kicks Editable / Deconstruct Kick Samples скачать в хорошем качестве

How does Kick Analysis work? - Make Kicks Editable / Deconstruct Kick Samples

Just released KickDrum 2.7 and I thought it would be nice to share how these analyze functions are working (at least the one in KickDrum – can only speak for that) It's been over a year now since the first version of the analyze function was released. Back then it worked by making a zero crossing (ZC) map and creating a pitch curve from that – which was already great for a big portion of Kicks, but had some downsides: the biggest one was to isolate ZCs when a noisy texture was sitting on top of the sine (or quasi sine), which can create little mini zero crossings left and right of the ones we actually want to look at. For a human to see the fundamental is easy – we see right away and can say – this belongs, this not – algorithm wise it can quickly become edge case wild west. So in the first version there was a forward backward RMS path going over the signal: 1. first marking all samples that exceed the RMS line 2. for each of these peak areas search their ZC pair 3. mapping all the tiny leftover ZC, whose in-between amp peaks were not exceeding the RMS line: these were considered noise – and averaged out / or merged into their surrounding left and right RMS exceeding ZC pairs. 4. now building the pitch curve This was not sufficient and was breaking on too many edge cases. A click sample with noise that is loud in the transient could exceed the RMS and then its ZC pairs would be collected, even though it's not matching what a human ear and eye would consider correct – it was not following along. Version 2: The second version tried a different approach: we solely looked at the spectrum (frequency domain), which is less deceptive in terms of pitch, but also has the downside that we lose time domain resolution – bad time domain resolution? let's just throw more compute at it! But offering more compute first makes it slow, secondly still does not give enough resolution in time domain: Because ideally we do not only want the pitch curve, but also we want the zero crossings, because the re-oscillated sine wave compound (fancy term for kick) should not only match in pitch, but also share the zero crossings, so that the analyzed version matches the phase. The strength of this approach was that it was robust against noisy zero crossings, but was slow and had difficulties to capture fast changing pitches precisely due to lack of time domain resolution. From these two versions I learned that the solution must lie in the combination of the two, since building the pitch from zero crossings is the best, while the spectrum was robust against noise. Version 3: With this in mind, I built a little playground where I could chain together different algorithms. After two days of playing with it – the solution was given: We break the signal apart into a parallel series of filtered bands (multiband) – looking at these tells us what frequency band, when, is the loudest, telling us: okay, the fundamental pitch of this signal goes at this time in this band and at this time in this band, judged by their gain envelope alone. All these put together give a pitch curve – it's a rough approximation, but what follows got me really excited: If we now take this approximate pitch curve and perform a bandpass filter sweep over it, with a bandwidth wide enough to compensate for the real pitch deviation from the approximation and tight enough to keep out non-belonging harmonics, while keeping the center frequency from sample to sample on the approximated pitch, this should give us a filtered signal that exactly matches the input kick’s pitch fundamental, with all noise that makes the ZC analysis so hard completely filtered out – a clean, mostly sine signal, whose pitch is the fundamental pitch of the kick at each time interval: Now taking the ZC map and building the pitch curve is like a walk in the park – nice zero crossings everywhere. https://audija.com