Cherry Audio Forums

Hi,

really beautiful and useful module.
suggestion: (quantized) poly output?

thanks tom

Thanks!

Polyphonic pitch detection (like Melodyne's) isn't something a lot of people have figured out, and I'm not aware of anyone having figured out (good) real-time polyphonic pitch detection - I definitely haven't Of course, I haven't tried either. I can vaguely picture a machine learning kind of approach, though the scale of the training would be a bit much. In the more narrow case where the polyphonic signal were a combination of classic synth waveforms, perhaps the problem becomes a lot simpler. Maybe not especially useful though? Definitely something to keep in the back of the mind until a good idea strikes...

playertron wrote: ↑Thu Feb 09, 2023 3:24 am Polyphonic pitch detection (like Melodyne's) isn't something a lot of people have figured out, and I'm not aware of anyone having figured out (good) real-time polyphonic pitch detection - I definitely haven't Of course, I haven't tried either. I can vaguely picture a machine learning kind of approach, though the scale of the training would be a bit much. In the more narrow case where the polyphonic signal were a combination of classic synth waveforms, perhaps the problem becomes a lot simpler. Maybe not especially useful though? Definitely something to keep in the back of the mind until a good idea strikes...

Electro-Harmonix seem to have cracked that particular problem. With the '9 Series' there are many pedals which can decipher what each individual string is doing in a standard monaural guitar signal.

Don't ask me how they do it! I suspect it involves unicorn tears and phoenix feathers.

SYNTH9 demo on YouTube

I'm not sure how many mythical creatures were harmed by Electro-Harmonix as I'm finding it diffiicult to discern from a quick glance at that Synth 9 demo whether it actually is doing polyphonic analysis. The backing tracks are so loud and there's significant amounts of dry signal too.

It sounds to me like the polyphonic synth aspects are actually doing monophonic tracking but generating some fixed chord sounds that give an impression of polyphonic tracking. Although I could well be wrong.

Hi everyone,

I imagined something more like getting several voices from the sample (just like getting one voice now), or something like connecting several of these modules with the same sample (but different settings) to create a stream of voices, that would be additionally quantized according to some scheme, or the user himself quantized them as needed. I hope you understand what I mean. Rather, several identical modules in one.

thanks for attention tom

basa333 wrote: ↑Thu Feb 09, 2023 2:35 pm I imagined something more like getting several voices from the sample (just like getting one voice now), or something like connecting several of these modules with the same sample (but different settings) to create a stream of voices, that would be additionally quantized according to some scheme, or the user himself quantized them as needed. I hope you understand what I mean. Rather, several identical modules in one.

Tracking several different pitches within one mixed signal is an entirely different proposal to deriving a voltage from a single, monophonic signal.

When you only have one pitch to track (monophonic operation), you can use simple techniques such as a Phase-Locked Loop or Frequency Counter to extract the pitch and drive a voltage controlled oscillator.

That's why old-school guitar synths used to have hexaphonic pickups. These would isolate the signal from each guitar string in turn, so that it could be tracked in a monophonic manner. In the synthesizer unit, you'd have six parallel monophonic voices, each containing:

1. A 'simple' monophonic pitch-to-CV extractor (PLL, Frequency Counter, etc)
2. An envelope follower
3. Voltage Controlled Oscillator
4. Voltage Controlled Filter
5. Voltage Controlled Amplifier
6. (Optional) an ADSR envelope generator, triggered by the Envelope Follower going over a certain threshold
7. (Optional) an LFO for modulation effects

Modern units, which can track multiple pitches within one mixed signal, rely on complex DSP techniques to 'unpick' the individual pitches. I suspect that those techniques would require a LOT of research, or which are simply beyond most of the VM developers. It's certainly something waaaaay above me!

Thank you!

Many greetings tom

Might have to get that EH thing and try putting it through some tests! But yeah, a lot of things have kinda cool yet fake ways to get poly info from poly audio. Nonetheless, I do have some ideas for tackling the problem one instrument at a time - knowing the audio source can go quite a ways to some preprocessing and assumptions that can give you a pretty good head start for something like guitar chords. Been tinkering with some ML methods too, which also improve with some simplifying assumptions. Right now though, this is more of a long term background project - the work/benefit ratio isn't great and learning a midi controller isn't that hard Working from recorded songs seems like a distant holy grail better left to a full team of engineers with a good deal of resources...

Here's something I'm curious about: Sending the pitch output to a 1v/oct cutoff freq on a steep filter - audio sources detected by the first unit can then go through the filter to a 2nd pitch detection, presumably with some initial note largely removed so that a 2nd note can be detected... and you could go on like that a couple times. I can think of at least a few reasons why this won't actually work, but it sure would be interesting to see what it does.

Hmm, that's an interesting idea - doing successive adaptive filtering and processing.

If I had to attack polyphonic tracking then I'd do a fast Fourier transform, scan the frequency domain results looking for the highest three or four peaks then do some semi-intelligent quantization to end up with the nearest half-sensible triad.

Possible improvements would be to apply carefully calibrated low-pass filtering and compression prior to the FFT and maybe do parallel monophonic tracking and use that pitch info to tweak the quantization in order to cope with things like pitch bend and vibrato.