VSTI world hate MIDI GUITAR?

[MaxJoy]

I did some experiments with feedbacked pitch to control the cutoff frequency of the input filters. The hope was to get a more stable pitch after the wild init phase. But notes still tend to jump to the next octave :-/ .

AQ: Please spend those few hours to make some practical expierences with pitch2midi. This is a real biest, and till now FS's approach is the best approach I've seen so far.
Sorry for the trouble you're in. Hope I didn't insult you.

[MaxJoy]

Did anybody ever consider FFT to analyze the pitch? At least some impuls response reverbs provide zero latency nowadays, so maybe it's possible to keep pace with the zero-crossing pitch tracking method.



With more frequency informations it would be possible to analyze the whole sound spectrum of the swinging string. For example you could distinguish an E on the 5th fret of the B string from the open E string. Or even analyze chords with todays computer power.

But maybe I'm just dreaming ...

[MaxJoy]

Some more thoughts: The human ear is able to recognize a sound even if it's damaged in many different ways (missing bass frequence, holes, noise, concurring sounds etc.).



I think this is a holographic trick: Instead of focussing on only one frequency we know the fingerprints of many sounds and are able to reconstruct the whole sound from a little piece of texture.



Maybe a job for a neuronal network?

[AdmiralQuality]

Did anybody ever consider FFT to analyze the pitch? At least some impuls response reverbs provide zero latency nowadays, so maybe it's possible to keep pace with the zero-crossing pitch tracking method.

I don't think FFT has narrow enough bands to get anywhere near an exact pitch out of it. If it did, you'd have thousands of bands (preferably 10s of thousands) and you'd have to lose the first F off of FFT. Also, if you *could* do this, I'd want to have a go at figuring out CHORDS with it. Thought about this already more than a little, but yeah... holy grail.

With more frequency informations it would be possible to analyze the whole sound spectrum of the swinging string. For example you could distinguish an E on the 5th fret of the B string from the open E string. Or even analyze chords with todays computer power.

But maybe I'm just dreaming ...

Well remember that the whole harmomonic series on a guitar note changes with which pickup(s) you're hearing, as well as where on the length of the string it was picked. Not to mention that it kind of evolves over the length of the note. The Axon actually can sense a few different pick ranges and turn this into controller (or was it sounds on different channels? Forget.)

This is part of the problem with chords. Even one note looks like a chord because it kind of is. Everything can be thought of as a chord of sine waves.

But, those sine waves in a given single note are all whole number multiples of the fundamental (at least in clearly pitched tones, Gamelan music and you're in trouble ). So there should be some hope in correlating the mess of harmomics in a chord back to the fundamentals... telling which harmonic belongs to which note... and being able to detect fundamentals even with strong harmonics as you would just follow the series you detect down to it's natural fundamental. But again, even getting a precise description of what harmonics are going is as-yet unattainable. At least it certainly isn't in real-time.

[AdmiralQuality]

Some more thoughts: The human ear is able to recognize a sound even if it's damaged in many different ways (missing bass frequence, holes, noise, concurring sounds etc.).



I think this is a holographic trick: Instead of focussing on only one frequency we know the fingerprints of many sounds and are able to reconstruct the whole sound from a little piece of texture.



Maybe a job for a neuronal network?

Neural network is how the Axon claims to work. Their lit says they trained one with a bazillion guitar note attacks until it started to get accurate. Now, it's not like the input to this NN could be purely just the raw incoming samples... as then even volume changes would create an infinity of possible attack samples. Phase of the harmonics against the sample rate is another biggie. (The exact same note, played exactly the same, sampled twice, won't come out as the same series of numbers.) So they're probably doing some kind of pre-processing to simplify or reduce the attack sample into some much smaller set of descriptive parameters, which they THEN run into a NN. What exactly, I can only guess. But I'd like to know.

Yes, our hearing is an INCREDIBLE processing system. Remove a fundamental, no problem, our ear will IMPLY it just because it's so good at exactly what I mentioned in the last post: correlating harmonic series to their fundamentals. This is why you can listen to a piano performance and transcribe it purely by ear (if you're really good).

Note that your ear doesn't sample. Actually it's closer to the ten thousand band FFT I described above. Cochlea in your inner ear is a long coiled up tube (looks like a snail) thick at one end, thin at the other, full of liquid, with thousands? millions? of little sensing hair-cells on the inside. So what your brain gets from that is like a massively parallel FFT coming at it. So like I described above, that's 90% of the problem. The math to figure out what harmonic belongs to what fundamental is fairly straight forward (though also becomes massive when there's a large number of harmonics as each harmonic needs to be compared to all the others... problem increases exponentially as you add harmonics.)

Picking out a voice in a crowded room is an incredible task. Even hearing the words the singer is singing over piano. Probably things computers won't be able to do for some time still.

[MaxJoy]

Ok, let's simplify the problem:









- Forget chords, at least for now .








- Limit the frequency range: The first 3 harmonics (base note, octave and 5th) are enough to identify a guitar note, this is the minimum fingerprint we need. Together with the four octaves a guitar can produce we sum up to ca. six octaves that we have to analyze.





- The minimum note difference we can hear is around 3 cents (3/100th of a halfnote). This is the FFT band width we need. 6 octaves divided by 3 cents = 2400 bands. Well, at least better than 10 thousands of them .





Supposed this considerations are correct, the quality of this algorithm depends on the intelligent post-processing.

[AdmiralQuality]

Ok, let's simplify the problem:
- Forget chords, at least for now .

Oh don't worry. Just not gonna happen. If I could do chords I'd make 10 million dollars in a week.

- Limit the frequency range: The first 3 harmonics (base note, octave and 5th) are enough to identify a guitar note, this is the minimum fingerprint we need. Together with the four octaves a guitar can produce we sum up to ca. six octaves that we have to analyze.

Six octaves, with 12 notes each, with lets say 50 cents of pitch offset for each semitone (we'll allow a little "slack" in our tuning precision, it would actually be 100 cents to allow for +/- 50 in both directions)... so that's an FFT with 15,000 bands.

No such thing.

ALSO... FFT implies some latency for its window and that's exactly what we're trying to avoid.

I have some other ideas. And again, apparently some of these guys don't mind playing on the neck pickup only, having weird pitch artifacts, and some of the other problems with FS's tracking system... they think it's a feature! So all I'm offering is to do the same "naive method" - i.e. timing zero crossings... but I'll spend a BIT of time seeing if I can't improve it a bit... through filtering... and.. a few other ideas that are too much to get into here.

- The minimum note difference we can hear is around 3 cents (3/100th of a halfnote). This is the FFT band width we need. 6 octaves divided by 3 cents = 2400 bands. Well, at least better than 10 thousands of them .


Supposed this considerations are correct, the quality of this algorithm depends on the intelligent post-processing.

Noooooooo... cents arent 1/100th of an octave. They're 1/100th of a semitone! To do 1/3 cents... it'd be

6 * 12 * 300 = 21,600 bands.

And anyway, even 2400 bands is un-doable.

[MaxJoy]

For me 6 * 12 / 0.03 give 2400. Anyway, too much it seems . Ok, let's wait another 20 years .

[FrettedSynth]

So you think it could be done workably? I think you could be the prickliest of characters, and still sell this device. It's what I was looking for before I started looking at midi guitar. Do you have any ideas as to how to implement a pitch to CV or pitch to midi in Synth Maker? I'd ask Fretted Synth, but I don't think he'll coming back to this thread(understandably).
Cheers, Nix

Couldn't stay away, my head said Yes and my heart said NO Not much to say to AQ, he is obviously a smart dude and I wish him well in his endeavors.
I notice that others are interested in guitar synths as am I, That is my biggest reason for making these in the first place, nobody was! so I did it myself. I have learned a lot along the way, considering these started in about 2001-2002 in Sync modular.

Just so you know where I come from a bit! I'm not an engineer or a DSP guy heck my business don't even use computers in its operation. I'm a bass player. My teachings in music\gear stem from? Being the guy in the 70's to early 80's who had to make sure the system was running and sounded ummm? I think OK?
Through the late 80's to mid 90's I became a gear addict "Yes I admit I am an addict" Then in about 95 I got my Mac IIci rushed out and bought a audiomedia II card bundled with sounddesigner II and a copy of deck, Man I had it all for that time haven't recorded much since then, The computer sucked me in

Modular synth programs started for me in about 98-99 can't remember? Sync modular was at v0.59 if I remember right? Got the free update to Reaktor and tried many others along the way (still love them all). Synthedit is the one I have used most lately, Why? because I feel comfortable using it and feel Jeff has done a great job, along with the many others who have created SE modules.
The question of "Why Free" ? Time is the biggest reason, My daughters and family are who I need to spend time with, so I could never give support the way a true business should IMHO. Do I think its because I use SE that I could not provide that support? NO. I have fixed any bugs that was reported to me! as any good business should, I take that back for some reason FreeAmp2 won't run on linux and have a problem with a couple of the VST adapters but I will figure that out also. To my knowledge all my other plugins have NO bugs even on linux?

Many have asked? why not put up a donation link? and I thank them people for the offer (not that I don't need the money) but IMHO most who would donate to such a link are! Kind hearted, hard working middle class people (not unlike myself) To them I have to say "keep your money and take care of your family".
They are not the people of the world I want to profit from, if I profit from anyone at all.

Then the question of SE and other freeware developers (dare I say that word "developers") clouding, convoluting or blurring the commercial market. I have to agree with what has already been said " if its a truly innovative product it will sell" Just the anti has been upped a bit by the many offering out there now. Before you can say "professional" (which I hate that word) you best be sure it really can deliver.

BTW I should have mentioned earlier "PhazOsc" is my newest synth. The rest are in serious need of updates as they are a couple years old, The bass synth is almost done except the UI (Rick you out there)

Anyhow enough about me! lets talk "Guitar synths"

Peace
Fretted Synth

[AdmiralQuality]

For me 6 * 12 * 0.03 give 2400. Anyway, too much it seems . Ok, let's wait another 20 years .

Nope. 6 octaves times 12 semitones times 100 cents. As you said 1/3 of a cent. So there's 300 of them in a semitone.

Anyway, even 2400 is too many. Forget the processing the FFT would have to do... just handling the *output* of that function would swamp your CPU. 2400 bands at how many times a second to analyze them? And here again it depends on the window size of your FFT. Aside from inherent latency, FFTs usually work on a brief snapshot of a chunk of time, rather than a running total with finer granularity. Wouldn't be very well suited to providing smooth controller streams

Actually though, it would be interesting to hook up an FFT of a small, reasonable, number of bands, and use it to produce a few MIDI Control Change message streams. You'd never get useable pitch control out of it, but it could be a frequency sensitive envelope follower. Kind of a CC vocoder!

[MaxJoy]

I never talked of 1/3 of a cent. 3 cents! Just my 2 cents .

[AdmiralQuality]

I never talked of 1/3 of a cent. 3 cents! Just my 2 cents .

Oop! No you didn't. Well OK then... there are 33.33333333... 3 cent increments between halfsteps then. Not 3.



By the way, that 3 cent thing is just what you can hear if you're playing the note in isolation. Play the note along with another note, in a chord or in a song, and I'd say the ear's sensitivity is more like 1/3 of a cent. Cuz you hear the beating.

[MaxJoy]

Most people don't hear the 2 cent difference of our equal tempered fifths. Not to mention the totally mistuned thirds and sixths ...

[AdmiralQuality]

Well they hear them if you play in some other scale with perfect intervals. At least they hear the difference. They just EXPECT these "mistunings" in western music.

I know I hear immediately if it's a real Hammond (or a proper sim of one) vs. a sampler. Due to the tuning errors on the real/sim Hammonds. One note at a time... maybe not so much. But as soon as the chords come out I know.

I'm playing with my synth right now... 3 cents difference between oscillators is quite detuned sounding. For middle C it's about one beat every 2 seconds. Not totally warbling by any means. But it certainly sounds "chorused".

Anyway, you're right, probably as close as any guitar player can play.

Cheers!

[guitarzan]

lets talk "Guitar synths"

Amen to that, brother!
My only reason for bringing up your synths in this thread was the hope that some would see there is hope beyond MIDI for guitar synths. Twenty years or so into MIDI guitar and it still isn't really doing it for most! Not all advancement is going to come from new code - thinking out of the box is what is really needed in many cases.

That's why I brought up another new development in the same post where the whole pitch2CV argument got started, the Electro Harmonix HOG. Now, I'm not pushing these things - I've never even seen one - and I know that they are probably NOT really tracking, probably no oscillators - BUT, they are apparently identifying pitch well enough on all six strings at once to pitch shift/double every one of them without using a hex pickup - and no glitches (they claim). It seems to me that if that is possible, then you could get polyphonic guitar synth without a hex pickup?