Hello,

I am a sample library developer (www.soundemote.com) and also do work for various sample library companies. I am looking for a way to take a sample, say a guitar pluck, and stretch/compress the audio so that every cycle of the guitar pluck wave is exactly synced with a sine wave, say at 440hz (A4).

Essentially I want to phase lock an imperfect acoustic tone to a perfect digital tone.

To start simple, let's just think of the guitar pluck as just another sine wave, but not perfectly in tune. I ultimately plan on having the guitar pluck reduced to its fundamental for the calculation of stretching/compressing.

Another way of thinking about this: Take a sine wave sample at 440hz swinging from +1 to -1 amplitude. Every 2nd time it crosses 0, that is a cycle. Take another sine wave, say 441hz. Cut that up into its individual cycles, then reassemble them so that each piece is compressed to match the length of the cycle of the 440hz sine. That's what I want to do to acoustic samples

If this could be accomplished, it will make a lot of people happy and I'm wondering if PD might be the right software to accomplish this.

Is it possible with PD? And... where do I start?

To an extent this is possible with existing Pd objects. I've tried a similar thing for a live sampling tool, but found that there are some percussive sound types which are hard to classify: glass objects, claves, bells. Pitch detection for such objects may be unreliable when using standard solutions to find periodicity. Therefore it is hard to make it 100% reliable. But in your case you process files offline, so you can easily discard results which seem to be wrong.

Anyway the method would be to first detect pitch, using [fiddle~], [sigmund~], or [helmholtz~] (http://www.katjaas.nl/helmholtz/helmholtz.html). Then calculate the required amount of stretch, and resample using [tabread4~]. If the sound contains high frequencies above SR/8, you can get aliasing. If aliases are noticeable it is better to first filter out those frequencies using one of the high order filters in iemlib. Alternatively you can do resampling with a command line tool like SoundTouch (http://www.surina.net/soundtouch/). Then, to synchronize phase with your sine wave you must first identify the periodic part (exclude the transient region) and apply cross correlation to find the phase offset.

Katja

Edit:

Crap, it seems that I did not understand the idea. The part about cutting up the wave in periods and reassemble is not clear to me, so my reply makes no sense at all. Sorry.

Another attempt, let's see if I get the point this time.

To cut a wave in periods and time-shift these periods is essentially PSOLA pitch shifting / time stretching. There's a lot of texts about that online. I've recently experimented with this technique for real time pitch shifting with formant preservation.

It works basically like this: you detect pitch in the signal with one of the standard methods (I used [helmholtz~]), then cut the signal in lengths of two times the period lengths, window these segments with a Hann window and shift them in time according to the desired pitch. Within a limited range, this works remarkably well. If you go down one octave, the segments do no longer overlap and it sounds rather artificial for voice. But for some instruments it is still okay. If you go up, the windows sum to a constant, so no amplitude modulation. Formant spectrum is not modified, only pitch.

For real time processing like I did, the segment repositioning can not be done with regular Pd objects, but for offline processing you may be able to do it partly at message rate with regular objects. Also you would have the opportunity to identify 'pitch marks' to improve results. Obviously you would not process transients and other non-periodic sounds in this way.

Katja

I don't want to cut a wave per say, that is just my example of "how one might think about the process." I don't care what process is used to get to the final goal. The final goal is simply "have the period of an acoustic wave sample to be equal lengths throughout" Actually, syncing to a sine wave is not necessary, the sine wave was there just to reference a musical note / frequency / period. I can manually tune the sample to any pitch once the period is fixed, but that's all I really need. Fixed period lengths.

Thank you for your explanation so far! Please continue to respond until we are on the same page. You might have already given a good solution, it's just a bit over my head right now.

I think what you're after is just reading a wavetable at different pitches but starting at the beginning at periodic cycles? if so this can be easily accomplished with loop~, tabread4~, and *~

I need to stretch/compress each period of the wave so that each cycle is equal length. Is this where you're going, sebfumaster?

I don't actually use PD right now, but I think it would be a great resource for denoising and tuning.

I am currently exploring a solution using REAPER and scripts. I was already able to do what I am after, using REAPER and manually editing the files with a text editor. All I need to do now is create scripts to manipulate the text, so that's what I'm working on now. REAPER can already find the 0 crossing and reduce samples to a fundamental. But the end result may or may not be good enough without further analysis of the original wav and smarter algorithms (PD might do a better job).

Maybe I need to post some before/after results, images, stuff like that so we are on the same page. Will get back to this thread.

So basically you want to get rid of vibrato? (Seems I missed the point two times already). Stretching / compressing periods using [tabread4~] is simple enough. Only thing is how to automatically identify start and end point of each period. Some sounds have more zero crossings in a period because of strong overtones. Therefore, after identifying average pitch you could use a filtered version of your sound file to see what the fundamental is doing exactly. I would recommend zero phase FIR filtering, implemented as fast convolution in frequency domain. Four times overlap and Hann windowing before fft and after ifft, you can do very steep filtering. Pd has all facilities you need.

Katja

Great, that's the first step, knowing it's possible. Thanks Katjav!

i think one big problem you'll run into if using 'zero crossings' on complex non-sinewave signals, is that the actual crossing points of those waveforms won't mean too much at all once you start dealing with anything other than very simple waveforms. Once you add all sorts of tonal colour to a waveform, then the actual 'zero' point will only be as accurate as the tonal purity of the original waveform.

i mean, look at this example, of a fairly pure tone, that of a plucked guitar string:

http://i.stack.imgur.com/4qAjK.gif

even that, the zero crossing are all over the place. So as Katja said, you'll need to pull the fundamental out and then somehow map that back onto the original waveform's period.

I think it all starts to get pretty intense quickly if you're not a great mathematician.

Yes, that is why it is importable to use an FFT to isolate the fundamental, and my tests have proven my technique requires tools that are already available, no math needed, at least not from my brain, because I am horrible at math compared to what would be needed to write code for PD. But surely the modules already created by others will be sufficient. Because I run into ridiculously stupid problems using the tools created by DSP companies, I will be exploring PD in the future for tuning and denoising. Based on what I know using the software I use, I bet the PD modules can do better just because I will be able to "have more say" (more flexibility) in what PD does.

There are no tools for sample processing, and it is incredibly frustrating. I have to use tools meant for vocals and speech. I think with PD I can finally create tools specifically for sampling.

What I think is, you can brick-wall filter just above the average fundamental as found by a pitch detector. One octave above it or so. If you do this with a zero phase fast convolution FIR filter, it is easy to implement and also it has a known constant time delay which is equal for all frequencies. Therefore it is easy to map to the original file.

Katja

Edit: see http://www.katjaas.nl/FFTwindow/FFTwindow&filtering.html for an illustration of how and why it is possible to brick-wall filter in frequency domain with very low level artifacts, if only you use sufficient windowing. Actually, the windows in time domain smoothen the brick-wall in frequency domain so it is not truly brick-wall filtering but it comes close.

I am continuing this thread under a new name: What is the best way to isolate FFT frequency bands?