G'day Daniel!

Yes very busy. Sorry for missing this post.

Yes I remember now, the idea we both had simultaneously for a new option in the FTS 3.2 beta to set the channels to play on each device, instead of parts, to use with MTS sysexes so you have a different sysex for each channel, to get an amazing amount of "pitch polyphony", up to 2048 unique pitches on one midi device if it supports per channel MTS sysex retuning.

I'm not sure I understand rightly. It sounds as though you are saying that the individual voices on the Yamaha SW1000XG are actually in different keys to each other when you play the same midi notes through them, so e.g. a 60 on one voice will play a C and on another voice the same midi will play e.g. a C#. But that seems hard to believe as it would make the soundcard unusable for many purposes. So I think I must have misunderstood in some way. I could believe that they vary by a few cents. Perhaps you mean the actual samples used for the voices in the wave tables?  - so that e.g. one voice would be tuned to a 60 at the correct frequency for middle C, another tuned to 61 for correct pitch for C# etc.



Generally wavetable synths use different audio samples for each octave or so. It's to do with the timbre of the note, an iinstrument may sound very different in different octaves and just doubling the speed of a sample may not sound like the same instrument played an octave higher.

Within an octave they may use only one sample for entire octave, or only a few samples, depending on amount of memory needed.

Ideally you would want different samples for each semitone since each note has its own individual timbre on most instruments. Also you would want different samples also for many different volume levels for each semitone, as the quality of sound on an instrument varies as you change the volume so e.g. a note played ppp doesn't sound the same as a note recorded as FFF and then played back very quietly.

So that's what the like of Gigasampler etc do, they sample each semitone, and sample it at several different volume levels requiring gigabytes of memory for each instrument, so streaming from disk. They actually get players of the instruments to play every single note on their instrument at many different volume levels to get the samples. But wavetable hardware synths don't have the memory for all those samples, and have much smaller numbers of samples.

So - anyway so you can't rely on the same note being the same pitch in different octaves to better than a few cents, on most wavetable synths. I would expect that you need to retune all 128 notes individually in FTS. Sometimes they may use the same sample for more than one octave in which case e.g. all the notes on the instrument may be sharp or flat by the same amount for two or more octaves.



Yes you should be able to do that, I think. When you open a .tun file in FTS then what it does is to convert it into a scale of 128 notes - so yes, you can retune every note in the .tun file individually in FTS - and then resave the result. You can also do that in real time as far as I can remember, in the Scale window. I'll take a look and say a bit more about that in a minute in a separate post.

Since they use the same sample, you might well find that all the notes in an octave, or even several octaves, are sharp or flat by exactly the same amount in cents, which may speed things up a bit.



Yes that should be possible by using the Scales for Parts. It's just a 128 note scale after all. You can also shift individual parts by any amount in cents in the Parts window (octave shift column, change droplist at the top to Interval).



Yes, that is an astonishing level of pitch resolution. But I can well believe that it is possible. A lot of it sometimes is just to let yourself  believe you can do it when others may say it is impossible, and learning e.g. to hear as a pitch difference what may originally sound like a difference of volume (because notes which are sharper by tiny amounts such as a cent or so or less, often just sound as if it's the same note slightly louder - or quieter - depending on whether the original note is high or low in pitch).

It's like looking at really faint galaxies in the night sky. Astronomers have the same eyes as everyone else, but because they spend ages looking at the night sky they can see faint smudges that are invisible to anyone else. Maybe partly eye gets more sensitive and maybe partly you learn to notice differences that most observers would find falls below the threshold of perception as "noise".

I know myself that I can train myself to hear finer and finer pitch discriminations. But nothing like that. For most musicians it is good pitch discrimination to get better than 4 cents (1 Hz at about middle C) for isolated individual notes and 10 cents, a tenth of a semitone is nothing to be ashamed of as a musician. You get much better discrimination for notes in chords if you listen out for the beats as a piano tuner does. But 0.02 cents means you are doing better than any tuner could do. At that level then you are talking about minutes for each beat.

So, obviously not using beats, unless you need to listen to each note for several minutes to do the discrimination, must be using direct perception of the pitch of the notes at that level of accuracy. If it is like the astronomers, then you probably don't have significantly more "hairs" in your ears to perceive the notes, just learnt to be able to pick out a slight variation in the signal from them that most people don't notice at all, is my guess.



With the wave count method - you can get it to show the wave crossings, then can inspect by eye to see if it looks as if it is counting the waves accurately. It could show a variation of pitch by  multiples of 1 Hz if it skips a wave by mistake which it could do with a complex waveform. Or if the waveform is inharmonic, which is quite possible perhaps for the Rhodes piano, it could get confused about where the wave crossings are since the basic frequency of the note could have another frequency superimposed on it gradually drifting.

So for a complex waveform, especially if it is slightly inharmonic - and for the most accurate measurements like this, I'd be inclined to go by the fequency spectrum rather than the wave count. You can tell by looking at the waveform as it plays. If the waveform is continually changing shape -well you get some change of shape during the attack of the note - but during the sustained part of a held note - if the wave continually changes shape then it has some inharmonicity. For a wave like that then the frequency spectrum is likely to be more accurate for exact pitch detection since the wave count method can get confused about where to locate the wave crossiing.

With the wave count then to make it as accurate as you can try really long notes like several minutes or more maybe ten, twenty minutes or longer though that mightn't be possible if the instrument decays. Well same applies to frequency detection method as well, the longer the notes the more accurate the pitch detection. But expect you already know that.

Both methods are a bit of an art rather than a science when it comes to finding the pitch of inharmonic timbres.

No don't have the soundcard. But sounds like it is worthwhile looking out for it on ebay. I wonder if it works with Windows 7 which is what I mostly use nowadays.

Thanks for your interesting questions.

I'll check out a few more details and then do a second post. Forgot to answer about the maximum pitch sensitivity for MTS, will need to look that up.

Robert



Linkback: http://robertinventor.com/smf/index.php?topic=58.msg132#msg132

Hi Daniel



Okay. That's quite normal for a wavetable synth I believe. With sound fonts, which perhaps are similar, you have to enter the frequency for each sample. I don't know for sure what the reason for this is - the technology and software is surely capable of much better.

After recording the note, the midi instrument designer must have to decide what pitch the sample is, when they set up the synth to play it, which may well be a cent or so away from concert pitch one way or another even if well tuned. I'd guess that very probably the entire instrument could easily be a cent or two sharp or flat of concert pitch in case of a keyboard instrument and with smaller variations from note to note.

So, perhaps it's just because designers of midi instrument don't bother to try to find its frequency to better than a cent or two, which for most midi applications is quite adequate but no good for the more demanding microtonalists. Not so easy to find software that lets you find the frequency of short notes to better than a cent or two, which is perhaps partly why it happens.



Yes been there myself.

Yes it's really hard when you get to the finest pitch resolutions to tell whether you can really hear a difference or not. I think that when you get to the limits, then you can often hear pitch differences but can't articulate them and e.g. you may know the notes are different in pitch but not sure which is the higher in the two. That definitely can happen when pitches are really close together and you are training in pitch discrimination - statistically it turns up random if you try to test for pitch discrimination, but you do actually hear a difference, just not yet got very good at hearing which of the two is the higher in pitch.

Also, that's very hard to distinguish from the situation where you think they are different in pitch but they are just slightly different in volume instead, because one listens so hard the mind begins to hear things that aren't really there.

You could use FTS to test for this - except trying it right now in the beta it doesn't seem to be working, so may be some bug to fix there.

Idea is to get it to use the random chord quiz in window 28, do them as sequences, set the scale to e.g. 1 cent (or whatever level of pitch discrimination you want to check). Set up three arpeggios, one to play the sequence 0 0 (no change in pitch), one to play 0 1, and one to play 1 0. So - objective is to see if you can tell whether the pitch goes up, goes down, or doesn't change at all. It's a sign of being able to tell the difference if you can tell that it changes pitch even if not sure which way it goes, so if you can tell the 0 0 sequence from the others, get it right more than a third of the times it gets played - then you can discriminate the pitch although maybe not sure yet which way it goes. You could expect then to learn to discriminate up from down for these very fine pitch distinctions with more practice.

I'll give that a go some time when I get back to work on FTS 3.2. Rather a lot on my "to do" list right now so not sure when I'll be able to work on the beta in a sustained way. Hopefully some time late spring or summer. But from time to time do a session or so of programming on FTS 3.2. so can give this a go.



I make that 200/8192 = 0.024 cents. But whatever it is plenty fine enough for even very demanding microtonalists, it would be a rare situation indeed where you need better pitch resolution than this.



When I read this, first I wondered if perhaps it is discarding the fine tune part of the pitch bend - but seems it is doing a bit better than that.

If you use just the coarse pitch bend range which is 128 (2^7) for the entire pitch bend range instead of 16384, it's 64 pitches to cover two semitiones instead of 8192, or 200/64= 3.125 cents resolution. Designers of a synth could argue that that is better than most musicians pitch resolution, so if the pitch bends are designed mainly for pitch bending using a pitch bend wheel, it will sound like a smooth glide rather than separate pitches to most musician's ears - though a few will hear the pitch glide as a sequence of steps like a microtonal scale with 32 pitches per semitone.

But 0.95 of a cent is between the two, better than coarse pitch bend, worse than the high resolution that fine pitch bend retuning is capable of. But it could be that for some reason the synth just isn't programmed to work to better than a cent of resolution for the pitch bends.



Okay rather sounds like an issue with the SW1000XG unfortunately



Oh the chorused piano may be a puzzle for it. It is usually done with several samples slightly out of tune with each other played simultaneously:

http://en.wikipedia.org/wiki/Chorus_effect

What happens with other voices?



Unfortunately, no. Is it the same for all the instruments? Try other instruments in case they differ. Especially since Scala also does the same thing, sounds to me like it may well be that the synth just doesn't do pitch bend retuning to better than about a cent for some reason. It's not likely that the midi pitch bends sent by FTS would be in error just at the cent level - most bugs are more obvious than this. Could be but a bit unlikely.

You can test to make sure. Try recording to a midi file as you play the two notes. Then inspect the midi file in the More (2 times) version of Record to File (Ctrl + 11). Go to List all events. Look for the pitch bends. It will say what exact pitch bends were sent.



The pb multiplier just multiplies the pitch before converting it to a pitch bend. It won't increase the pitch discrimination. Just means you can e.g. move a pitch bend wheel further to get the same amount of pitch bend.

To increase the resolution in midi the only way is to reduce the pitch bend range. You can make the pitch bend range 0.5 of a semitone and still play all the notes you want. If just 5 limit just intonation you may not need to retune as far as 0.5 of a semitone in which case you might be able to make do with a quarter of a semitone or better.

But that all assumes the device you are using has the capability of playing such fine pitch bends.

For really fine pitch resolution, then wave table synths are hard work because of the way they are designed around recordings of live musicians. While FM synths are much better in this respect because they make the sounds mathematically. Including the wave shape player instruments in FTS.

Though I understand, that's no solution if you want really good pitch control on natural sounding instruments.

I don't know if there is any tweak you can use for this - a question for any experts on how these hardware synths work perhaps.

Anyway yes - you can try out a midi recording in FTS just in case it is a bug and if it is of course I'll do what I can to fix it. But on the basis of what you've said so far, I rather suspect the issue is to do with the internal workings of the SW1000XG

Anyone reading this know any more about all this, about the pitch resolutions of hardware synths like this, how they work, or have any ideas about how to solve Daniel's question?



Thanks!

Yes, indeed completely agreed about that. Do hope I can improve this with FTS 3.2. I'm hopeful that I can since BM Pro has worked out well and I've learnt a lot from it. I just don't get the same reactions about BM Pro that I get with FTS 3.2. Its users generally don't seem to find the complexity daunting in anything like the same way that FTS users do. So I hope I can more or less completely solve this when it comes to FTS 3.2 when I have time to work on it properly.



Thanks, I'll let you know when I start work on it. I've got a short e-mail list of users who want to be notified when I start on the FTS 3.2. beta so will add you to it in case I forget.

Yes discussing it here on the forum seems like an excellent idea in case there are others who visit who can step in and help in one way or another. I'll do that when the time comes.


Linkback: http://robertinventor.com/smf/index.php?topic=58.msg134#msg134

So.... Robert,

I've done a little reading on XG to find that the pitch bend sensitivity is set to +- 2 by default. And it does indeed use the full 14-bit resolution so I do not know why the voices will not tune to a resolution of better than about 1 cent. I cannot get in between the gaps of about 1 cent!?!
 On the plus side it seems that the Spectrum Analyzer in FTS is very good. Repeated recordings of the same pitched note usually give readings varying of only 1,000th a hz, if that - so it should prove to be a useful diagnostic tool!

One thing that I found is that on the XG you can use an RPN message to change the coarse tuning or fine tuning of the note. Coarse in in semitones; fine is in 100/8192ths! - which should provide excellent resolution. I've yet to see if the fine tuning parameter will sufficiently change the frequency of a note but i can easily test that in xgworks tomorrow. I'm guessing sending an rpn will change the value for the whole channel not for a single midi note but..... a solution immediately sprang to mind.

Instead of sending pitch bend messages, you could send RPN fine tuning messages to seperate channels depending on the midi key pressed - just like the method in FTS for pitch bends.

Unfortunately, I have no idea how to do this - so thought i could ask you! It might be easier or harder than I think.....

What do you think? I think the idea has potential for a lot of people who want to retune their sample based synths to a really good resolution - if it works?!?

So I'll test the tuning of that fine tuning parameter tomorrow.   

Great - another crazy, ambitious, workaround. You know, I resent being a good-eared, caring, sensitive musician microtonalist guy I really do!!

So of course you will need to apply an offset for EACH midi note just yo put the damned wavetable back into perfect ET, and then you can either apply the pitch bend method, but that depends on whether the voice tunes to the full pb resolution (which it wasn't).

The other way is to have a tuning table full of all these offsets and the fine tuning rpns for each scale for each voice!

And another way is that you have an algorithm for the scale you have set in FTS to automatically add or subtract the needed fine tuning value for that scale....

But I don't even know how to send an rpn by pressing a midi key yet.....

If you like.... there are details on the rpn messages on the www.studio4all.com website and also by googling "xg guidebook".

BTW: Also of smaller concern is that when you load seperate .tun file for each part in fts, they will all tune to the freq in the pitch window. You have to aply an offset in the modulate by cents drop down. That's not too much trouble. It'd be nice to be able to have a feature to define the freq for each part, or even for it to load the .tun to the freq. given in the file.

It might be cool that when you modulate by interval (in the parts window), say 27/32 from C4, that instead of starting that scale from the C4 key again, you start it from the Eb4 key. In other words, being able to set the midi note for 1/1 for each part. Maybe I'll put that in the Wish List Section in the forum.

So before I get crazy on this idea, i'll check that the fine tuning works. But why would they have q fine tuning if it didn't work?? God knows, Yamaha?!?

Ok, seeya mate!

Thanks again.

Daniel

BBTW: setting these offsets makes a big difference, I can DEFINATELY hear it. If you like I'll do one for the SYXG50 VST so you can hear and compare, if you like??

Linkback: http://robertinventor.com/smf/index.php?topic=58.msg142#msg142

Hey robi,

Just giving you some more information for you about the SW1000XG. I have done some tests, and it seems that every 41 out of 8192 pitch bend increments will yield a change in pitch of approx. 1.15 cents. I checked this with the grand piano voice (I assume the other voices will follow suit). I checked by setting a pb in xgworks and recording it and then getting the fts spectrum analysis. I must have done this about 30 times! I have it in a spredsheet which i can put up here when i'm not on the iphone.
 So 8192 / 41 = 199.8 Which tells you that the synth has an intended resolution of 1 cent. Oh well. I haven't checked to see if changing the sensitivity to 1 will halve it but from a few preliminary tests - i don't think so.

So.... yeah, I've been a little bit miffed at yamaha for a few days. It definitely showed me that as a person that wants a just intonation scale with fairly good resolution (in the modern world), I've got my work cut out for me.

Oh well....... one mustn't give up. Maybe you know of an XG or otherwise synth that uses the full pb resolution? I might talk to Manuel about that...

You mentioned a way to set the frequency for soundfonts somewhere here on your forum. A good soundfont player is that sfz+ program plus you mentioned another one on another post.
Short of sampling my own fonts, how can i dictate the frequency for each sample in the bank? Will Vienna do this? You mentioned the autotune program. But that only retunes .wav files and the like doesn't it.

It seems like the next step for me would be to experiment in getting soundfonts tuned to exact ET and then use FTS to retune. I think many of the available soundfonts are probably out of tune from ET by varying amounts, would you say?

A quick way to detect their frequencies and then retune using some kind of FFT would be ideal. And then would they respond to the full 8192 steps or foes that depend on the player?

Probably my future is going to be in some kind of physical modelling, and using old analogue circuitry, (like the modular synths) to create great natural sounding microtuning synths with good resolution. I believe analogue synths could still create better resolution because they are on an electron rather than digital level.

Sorry, a lot of questions there. I will post those spreadsheets up soon...,

Daniel. 


Linkback: http://robertinventor.com/smf/index.php?topic=58.msg144#msg144

Hi Daniel,

Just been talking to Jeff Scott who is the software author of Li'll Miss Scale Oven for the Mac and knows a lot about the hardware side of things. He is much more expert in that area than I am.

So anyway - he had a few ideas, though none of them solve your problem, just to do with trying to understand why it might happen.

Anyway he also found it a bit of a puzzle. He agreed that it's unlikely to be anything to do with the actual calculations used to adjust the waveform itself to do the pitch bending, which should be way more accurate than is needed. It could be that the synth just discards some of the information in the pitch bends before processing them. Many synths use only 7 or even 6 bits for pitch bends sent when you adjust a pitch bend wheel  (128 or 64 separate values for entire range) - though with more recent synths 10 or 12 places are more likely (1024 or 4096 separate values). That's to do with the resolution of the analogue to digital converter from pitch bend wheel postion to digital midi pitch bend messages.

So even though the soundcard doesn't have a pitch bend wheel, its probably using the same chip, and it might be that they use the same resolution as the pitch bend analogue to digital converter (ADC) throughout. But if that is what it was, you'd expect the total number of pitch bend steps to be a power of 2 (e.g. +-64 or +-128 steps etc) which doesn't fit with your 41 pitch bend increments observation.

So - he thought from what we know so far, it might perhaps be due to the 12 note tuning system for the XG synth, which is adjusted in increments of 1 cent at a time for range of + - 64 cents. Maybe for some reason they use that throughout internally instead of pitch bends and transform pitch bends to the tuning in cents. Seems a bit surprising but may be the reason - maybe they designed the tuning table features of the chip first, and then added on the pitch bends later??

So anyway that is the best guess we have come up with so far, but not conclusive at all.

To know for sure need someone who can tell us how the hardware Yamaha XG midi synth processes pitch bends.
 
Wiki article:
http://en.wikipedia.org/wiki/Yamaha_XG

I found a data sheet for the Yamaha XG chips for notebooks here, which perhaps might be related
http://en.wikipedia.org/wiki/Yamaha_YMF7xx

But unfortunately, it doesn't tell you anything about how the midi synth pitch bends work.

However he also had a suggestion which may help you to find the resolution more quickly so you can e.g. try it out for many instruments, and varying pitch bend range etc, faster than using the FTS spectrum analyser or wave count to find the frequency of each note.

Also this could help to test new soft synths, hardware synths or soundcards as a quick way to check that they have something like the resolution you require (then if you find one that passes this test you can then measure the exact frequencies to make sure).

Perhaps you have done this already? If not then this may help.

The idea is rather than find out the exact frequency of each note, just look for when the frequency changes.

The easiest way to do it in FTS is to set it up with a scale of say 0.1 cents equal steps, so that successive notes on your midi keyboard or the on-screen PC keyboard player play pitches just a tenth of a cent apart. Just enter "0.1 cent" as the scale step instead of e.g. 100 cents for 12 equal.

Then play two note chords in that scale, keep one note as the 1/1, and then vary the pitch of the other note - and look out for sudden changes in the beats pattern, which will show that the pitch of one of the notes has changed. This will be easiest to hear with high frequencies.

If you have better than 1 cent resolution in your hearing for melody lines without chords, which I think some musicians do, you might be able to hear a 1 cent step. But with playing chords and listening for changes in the beat patterns, you can hear even tiny changes in pitch using this technique, well beyond whatever your best resolution is for monophonic lines.

That will give a first idea of what is happening. If you find all the notes are distinct in pitch with 0.1 cent steps you can try smaller steps e.g. 0.05 cents etc until you find out what the resolution is in cents (or you could use midi pitch bends similarly, just that doing this scale approach in FTS may be a quick way to do the experiment).

You can vary the pitch of the 1/1 to try this experiment for different frequencies. Then if it passes this test, you can then do frequency measurements with FTS as a final check.

So anyway thought that might be a useful idea to pass on.

Again, if anyone reading this knows anything about the design of the Yamaha XG synths that may bear on this do say. Jeff - do step in if you have anything to add or correct in what I have said.


Linkback: http://robertinventor.com/smf/index.php?topic=58.msg146#msg146

That was a real surprise, didn't expect that to happen with what is basically a soft synth.

Just tried it out with the Yamaha XG WDM soft synth and that does the same thing too, so I've duplicated your observations there - again that one has the jumps in pitch exactly at one cent intervals relative to concert pitch. Which has no connection at all to the way the pitch bends work with 4096 pitch bend increments to a semitone, it must be that it uses some other system internally based on cents rather than pitch bends.

Tried it out with the Roland VSC and that works exactly like the FTS triangle wave, really slow beats for the 0.1 cent intervals gradually increasing the further the notes are apart. So it has pitch resolution at least 0.1 cents, so that's an example of a wavetable soft synth with better than 0.1 cent resolution.

Should try out a few more but I don't have much of that sort of thing installed on my computers at present (mainly focussed on programming just now).

So anyway definitely something to do with how your Soundcard works I'd say,  and sadly, probably nothing you can do about it though if there is anyone who knows how the card works in techy detail maybe they could say more, so need to look for sample player, or sound font player or similar.

The Roland Sound canvas is a reasonable wavetable synth, not stellar perhaps but has nice sounds, comparable to the Yamaha soft synth, different in the actual sound set you get.

It's 5/4 is pretty much beatless on the Ocarina, maybe a little "soft" for high pitches but not bad at all. But on most other instruments the 5/4 on the Roland Sound Canvas beats quite strongly, with a wide variety of beat rates.

Sometimes it is because the instrument has a small amount of vibrato or tremolo when played on its own with no effects (usually meaning the recorded sample had level of vibrato or tremolo in it already when recorded) - but others I think may just be because the patches are a little out of tune - just now testing the 5/4 on the Roland Reed organ and it is beatless at some frequencies and has beats at others - not in a regular fashion it is just kind of random depending on both the choice of 1/1 for the 5/4 interval and choice of instrument - so I think that is a sign that the beats are caused by recordings in the wave table a cent or two above or below concert pitch, best guess anyway for why it happens.

So anyway if you had a wavetable synth, soft synth or sample player with the resolution of the Roland Sound Canvas - together with a sound font with perfectly in tune patches played without vibrato or pitch variation of any sort then you would be close to what you want.

If you did auto retune all the notes in the table for a sound font, if the individual notes vary in pitch slightly (not necessarily vibrato, just a small variation of pitch during the note) then you will still get beats when you play the notes with each other or even when you play unisons. So you need to make sure that the pitches are very steady for instruments of interest.

I haven't tried measuring the Roland Sound Canvas pitches with FTS but these tests probably give a good idea of what one would find if one did. I don't really have the time to do it at present as it is a bit slow measuring the frequencies like that - have to record a note, then take its spectrum and do the analysis for each one as you've discovered, and it needs to be automated more when I get the chance to work on it again.

The next obvious thing is to test sample players and the like. But I have a new installation of Win 7 on my computer which I've been using for programming rather than for composing or making realisations of tunes etc - so I haven't installed any of those yet. I'll give them a go some time.

Anyone reading this got any recommendations of wavetable synths, soft synths etc for beatless harmonies? Or any sound fonts that are already tuned sufficiently accurately e.g. all the samples accurate to within 0.1 cents of concert pitch (or preferrably better as you would still notice a 0.1 cent deviation as very slow beats), for this sort of work?


Linkback: http://robertinventor.com/smf/index.php?topic=58.msg148#msg148

Just fixed a bug in the FTS 3.0 release which may affect you if you try out the 0.1 cents scale idea from the PC keyboard instead of from a Midi keyboard

Details here: Tune Smithy 3 - Minor Bug Fixes - May 19th, 2010



Linkback: http://robertinventor.com/smf/index.php?topic=58.msg155#msg155