Physical Models

Dave Manley · Posted: Tue Mar 11, 2008 10:42 pm Post subject:

Perhaps a bit late, but the following query searches the US Patent Office for all patents with Assignee=Yamaha, with keywords "waveguide" and "dsp". It turned up a few interesting looking things. Note that some/all of these patents are still active.

shortened link

1 5,835,605 Full-Text Engine exhaust sound synthesizer
2 5,741,994 Full-Text Waveguide musical tone synthesizing apparatus with noise modulation of waveguide coupling
3 5,739,454 Full-Text Method and device for setting or selecting a tonal characteristic using segments of excitation mechanisms and structures
4 5,621,801 Full-Text Reverberation effect imparting system
5 5,519,167 Full-Text Musical tone synthesizing apparatus
6 5,508,469 Full-Text Musical tone synthesizing apparatus capable of changing musical parameters in real-time
7 5,438,156 Full-Text Wind type tone synthesizer adapted for simulating a conical resonance tube

Also: 5,212,334 Digital signal processing using closed waveguide networks

To get pdf's quickly use: http://www.pat2pdf.org/

The combination of searching on the USPTO website and pat2pdf makes quick work of digging through patents.

note blue hell: shortened the link to get a better page layout.

JovianPyx · Posted: Wed Mar 12, 2008 7:48 am Post subject:

Ah! Thanks Dave!

That should remove many question marks. I'm also interested in the "engine exhaust" synthesizer - I've used my (heavily modified) FatMan for that while flying with FS4... Confused

but I am very interested in how they do that. I really want to get this phLUTe working properly - as in "expressively playable" and this has a high probability of delivering answers to questions that loom.

jksuperstar · Posted: Wed Mar 12, 2008 8:41 am Post subject:

You can also use Google's Patent Search Engine, no need to a TIF reader, etc.

http://www.google.com/patents?hl=en

widdly · Posted: Wed Mar 12, 2008 1:53 pm Post subject:

I've played around with Perry Cook's STK a bit and the overblown flute sound is definitely my favorites. It's so heavy metal.

The models in the STK all seem to have some hard coded values for scaling the parameters. These scalers are never really explained so I imagine they are found stochastically.

You could write an application to systematically test the model with a whole range of parameters at different frequencies and automatically determine which combinations will actually generate sound. You could then use the results to generate key scaling parameters to extend the useful range of the model.

The model for the Steampipe in Reaktor is pretty interesting. There is a reverb inside the model. I don't know how that relates to physics but it sounds pretty awesome.

If you haven't tried it out, check out the freeware vst called Miles'tone. It blew me away Laughing

JovianPyx · Posted: Wed Mar 12, 2008 2:06 pm Post subject:

I am not yet finished fiddling around with this, but I have tweaked it to the point that it may simply be the inaccuracy of the fractional delays that is causing me grief. I mapped out the keys that work and the keys that don't and saw that the lowest octave plus a few extras all work, from there, random notes won't resonate and the ones that do resonate are fewer and fewer as I go up the scale. The way I see the fractional delay scheme, the fractional delay becomes a larger portion of the delay as the frequency increases. I think because of that, any error becomes more and more important.

I have an alternative that could be viable. If I map all notes back to the first octave, the mechanics work as they should. I can control the octave by decimation. 1:1 (or no decimation) for the first octave, 2:1 for the second octave, 4:1 for the 3rd and 8:1 for the 4th. This isn't exactly the way I planned or even wanted to do it, but I know I can make that work. Since this is a phLUTe and not a flute, I can get away with that Cool

. At 8:1, the system would take 105 clocks to calculate a sample. The sample rate of 100 KHz allows for 500 clocks per DAC enable so I could even go higher or I could get perhaps 4 pipes of polyphony out of it.

widdly · Posted: Wed Mar 12, 2008 3:33 pm Post subject:

That decimation idea is interesting. I would have thought you would need to filter at 1/2 nyquist before the decimation to avoid aliasing. I wonder if there will be a very obvious audible step as you change octaves.

JovianPyx · Posted: Wed Mar 12, 2008 3:45 pm Post subject:

JovianPyx · Posted: Wed Mar 12, 2008 7:37 pm Post subject:

Ok, it's sound file time. It's just a few notes of noodling in the lowest octave. There are things that need to be adjusted and messed with, but you'll hear the flute-ness right away. The clicks between notes are transients generated by changing the length of the waveguides, but I read on the Stanford site about a way to eliminate them. At this point, this is a sample of "first sound" from the synth in it's early stage of development. This hardware doesn't yet have the decimation logic, I'll post a better sound sample when the decimation and transient suppression is working.

phLUTe_001.wav

Description:

Sample of the phLUTe, just noodling in the lowest octave

Download (listen)

Filename:

phLUTe_001.wav

Filesize:

2.66 MB

Downloaded:

1180 Time(s)

widdly · Posted: Thu Mar 13, 2008 3:07 am Post subject:

The sample is sounding cool. The transients sound interesting, they have a string like boing to them. Maybe more Phlutey than flutey Wink

JovianPyx · Posted: Thu Mar 13, 2008 6:29 am Post subject:

It is possible that I am using the term "decimation" incorrectly.

Here is what I want to do:

The state machine which calculates the next sample runs very fast compared to the time between DAC updates (100 KHz). There is time to calculate more than 32 state machine samples in that span. When I say decimate 2:1, what I mean is that while the sample rate (to the DAC) remains constant at 100 KHz, I would calculate 2 samples (within one DAC sample time) using the state machine and use only the second (thus killing every other one or "decimating").

If the signal without decimation is a 250 Hz signal, then it is 500 Hz when presented to the DAC after decimation (but without changing the DAC update rate). At 8:1, this same 250 Hz signal is 2000 Hz. But again, the DAC update rate is constant. The upper end of the lowest octave has a top frequency of less than 522 Hz, so the output after decimation is less than 4176 Hz. Since the DAC sample rate is still 100 Khz and the signal conveyed is only 4176 Hz at worst, how does aliasing happen? AFAIK, the Nyquist frequency is 1/2 the frequency of the sample rate to the DAC. I see no difference between what I want to do and simply presenting a 4176 Hz sinewave to a 100 KHz sample rate DAC system.

widdly · Posted: Thu Mar 13, 2008 9:29 am Post subject:

Let's say you have 1 second of samples at 100Khz. So your state machine generates 100,000 samples. The highest frequency you can express in this block of samples is 50kHz. Lets imagine your samples contain a mix of a 5kHz sine wave and a 40kHZ sine wave.

Now you decimate by 2 by throwing away every second sample. So you have 50,000 samples. Now you have 1/2 a second of audio at 100 kHz and all the frequencies will have doubled. So your 5kHZ sine wave is now a 10kHZ sine wave but your 40kHz sine wave is now an 80kHZ sine wave. You cannot represent an 80kHZ sine wave using a 100kHZ sampling rate, so what you get aliasing.

JovianPyx · Posted: Thu Mar 13, 2008 9:41 am Post subject:

JovianPyx · Posted: Thu Mar 13, 2008 12:29 pm Post subject:

I think I see where the confusion is.

I have no idea what the technical term for this is:

When I am doing decimation by 2, I am running the state machine in an oversampled mode 2:1 and then I decimate it's output 1:2. This effectively doubles the frequency of the output signal without changing the sample rate.

widdly · Posted: Fri Mar 14, 2008 2:51 am Post subject:

It amounts to the same thing really. I think in theory the problem can still occur. In practice you said your bandwidth only goes up to about 5kHZ so it shouldn't be a problem.

You are assuming the bandwidth of the state machine is independent on the sample rate which I don't think it is. Think of it in terms of wavelengths instead of sample rates.

Lets say to generate a certain note, you delay line will be say 32 samples long. The highest overtone that can resonate in the delay line will have a wavelength of 2 samples.

To generate the same note in oversampled mode, your delay line will be 64 samples long, but the highest overtone that can resonate in the delay line will still have a wavelength of 2 samples. So effectively you have an octave more bandwidth in your delay line and when you decimate the signal, any overtones in this octave will alias.

JovianPyx · Posted: Fri Mar 14, 2008 5:33 am Post subject:

Tall ya what - I disagree, I think there will be no alias problem.

The proof will be in the sound file once I get the decimation scheme coded. That should be sometime this weekend.

widdly · Posted: Fri Mar 14, 2008 6:25 am Post subject:

Of course, sorry if have been too argumentative. I'll be interested to hear the results Smile

JovianPyx · Posted: Fri Mar 14, 2008 5:27 pm Post subject:

Here's another sound file. A chromatic scale starting at the top of the 4th octave down to middle C. I hear no aliasing. There are other problems, however.

phLUTe_002.wav

Description:

Full range chromatic scale

Download (listen)

Filename:

phLUTe_002.wav

Filesize:

1.44 MB

Downloaded:

1139 Time(s)

CJ Miller · Joined: Jan 07, 2007 Posts: 368 Location: 127.0.0.1

Three years ago I made some crude physical models on my OASYS PCI. Even these basic nodes showed a lot of potential. As I recall I gave up because my only controller was just note on/off so articulation was really limited. IAC sequences were too clumsy. I have too many old computers for different things! It'll be nice to try the OASYS again within my current setup once I can be bothered to get it out of storage.

As I recall i did some weird modal skin propagation nodes with additional delay lines between them. It was like a nonsense cross between tabla and saxophone, but without the controller data smoothed, and lots of scaling problems. But versions of this were my first success at making this sort of thing, it was noisy fun.

If you want to challenge your mind, check out the documents on physical modelling on Julius Orion Smith's web pages. I was interested to start on it, but I gave up after a week or so. Maybe I'll try to pick it up again later if I can fill in some fundamentals.

JovianPyx · Posted: Sat Mar 15, 2008 6:59 am Post subject:

JovianPyx · Posted: Tue Mar 18, 2008 1:44 pm Post subject:

A bit of progress today. In my reading of many many PDFs on this subject, there was some mention about the jet to bore length ratio. This had been expressed not as delay line length, but just bore and jet length. The ratio is to be set to 1:2 jet:bore. However, the block diagram in one of the PDFs shows 2 bore delay lines because that model was a flute with tone holes (much more complex). Anyway the model I'm using is a _single_ delay line bore model, it models the bore's normally 2 delay lines with one (derived from a similar situation in the Karplus-Strong string model). It occurred to me that perhaps the jet delay line should be exactly the same length as the bore delay line. The result was an additional octave of range, from middle C for two full octaves plus up to Eb.

So for now, the model has a minimally useful range which I will try to extend. The model is also rather hard to control in terms of overblow octave cracking. It cracks much more easily than I like.

jksuperstar · Posted: Tue Mar 18, 2008 9:47 pm Post subject:

Sample Rate and Bit Width can be related much like Voltage and Current. So, you can gain bits in the decimation of your rate, and so not loose any information, just transform it.