New oversampling ASIO driver

Scali · #1 04-06-2008, 10:40 AM

I bought a guitar processor a while ago (a Zoom G9.2tt to be exact), which has a USB connection with ASIO support. Very convenient for recording. However, it only supports 16-bit 48 KHz.
My own soundcard does 24-bit 96 KHz, and therefore whenever I wanted to add a guitar track, I either had to use an analog connection and record with my soundcard (which means plugging out other sources and plugging the guitar processor in), or convert my entire project down to 48 KHz or less, if I wanted to use the USB connection.
Or if I *really* wanted to have my hands full, I could make a copy of my 24/96 project, downsample the copy to 48 KHz, then record the guitar track... then import that 48 KHz track into the original 24/96 project, while having Cubase convert it.

So I thought it would be very convenient if the signal of the guitar processor could be upsampled to 96 KHz instead. This would allow me to preserve the quality of the other tracks, and as a bonus I could apply any effects in 24/96 on the guitar track aswell, which should improve the overall quality compared to doing everything in 48 KHz.. even if the source audio itself isn't of better quality.

Anyway, I started creating a proof-of-concept ASIO driver which can be placed in between Cubase and an actual ASIO device, and perform upsampling on the incoming signals, and downsampling on the outgoing signals (it does some basic linear filtering aswell, so the resulting quality is actually quite nice on the ears). Currently I have a driver which can double the samplerate, and I have it working for my guitar processor, and I've also done some small tests with my 24/96 soundcard, and it seems to work okay aswell. The nicest part is that there is no additional latency when using this driver. It costs a bit of extra CPU power, but on a modern machine it's negligible. The cost of having all VSTis, effects, mixing etc at 96 KHz instead of 48 KHz is far larger than the extra overhead that the driver adds.

If anyone is interested, I can clean the driver up a bit, shave off some of the rough edges, and release it to the public. I plan to release the sourcecode of this driver aswell, so other people can modify it to make it support other devices and perhaps other functionality.
It isn't a whole lot of extra work to add in support for more generic upsampling/downsampling... Currently you're limited to only the double samplerates of your device, but with a bit of extra work you could also turn eg 44.1 KHz into 96 KHz, or even 192 KHz or whatever you want.

Speaking of other functionality... Currently I can only use a single ASIO device for the multitracking. This means that I can use either my soundcard, or my guitar processor. I have this idea of adding multiplexing to the driver. With the resampling in place, I can basically make every ASIO device appear as having the same capabilities. Currently both my soundcard and guitar processor appear to Cubase as 96 KHz devices. The next step could be that I combine both to a single ASIO device with multiple inputs. This will allow me to record the guitar and other instruments at the same time, and multiple outputs could also come in handy, no doubt.
The current concept I have in mind, would give you an input latency equal to the highest latency device that you're using, and an output latency of twice the highest latency. In that case I think I can make it work in a robust way. With modern hardware, latencies of below 1 ms are doable, so if by using this driver you'd go up to 2 ms, that would probably still be acceptable. With older/higher latency devices, it may become a bit of a problem... You might also get problems when the latency between different devices is too large.
Anyway, currently I don't even know for sure if it will actually work properly in practice at all. But I am willing to give it a try.
I think the end result could be quite cool. If you were to combine the resampling, multiplexing, and the ASIO4All driver, then basically you can combine virtually any soundcard at virtually any sampling frequency and use them all in a single Cubase project at the same time. If it works well, it means you can create a fancy multichannel studio setup at home, by just combining a few cheap audio interfaces.

Anyway, feel free to express your thoughts on this, and perhaps you have other ideas that could be implemented.

axemanrio · #2 04-12-2008, 10:52 AM

Scali, you might want to give ASIO4ALL a whirl.

microdmitry · #3 04-14-2008, 01:42 AM

You do realize that your signal doesn't really become 24/96, right? Just like upsampling a 6 megapixel picture 200%, doesn't result in four times the resolution.

Scali · #4 04-14-2008, 01:08 PM

ASIO4ALL is not a real ASIO driver, and I take microdmitry's post as an insult. Ofcourse I know that, I have written the actual algorithm that converts the signal. What do you know?

microdmitry · #5 04-14-2008, 04:55 PM

It just seems mildly retarded to even use something that's 16/48. 16/48 is not enough for guitar signal. Upsampling it won't change this simple fact.

Scali · #6 04-14-2008, 05:57 PM

Care to elaborate? I'm not just going to take your word for it, seeing as I took a masters at university in signal processing and such. So we can discuss this in all the detail you like.
Feel free to bring friends like Nyquist, Shannon and Hartley into the conversation.

microdmitry · #7 04-15-2008, 01:50 AM

So did I. 48KHz isn't much of a problem (unless you want to do further nonlinear processing, which benefits from higher sample rates), but 16 bit is. 16 bit was created for pre-mastered, normalized signal that's at its ideal levels at all times. For distorted tone, you have a heavily compressed signal and this is much less of an issue. For cleans, however, they inevitably have to compress the dynamic range because there's not much of it available in 16 bit of resolution. Theoretical max is 96dB if memory serves, vs 144dB for 24 bit. Of course, 144dB is never realized due to the laws of physics (thermal noise in ADC/DAC resistors and preamp circuits, power supply noise, quantization errors, and such), but the beauty of it is that you have a margin to throw away, whereas with 96dB you don't. Interpolating to 24 bit from 16 doesn't change that.

That said, the signal in your modeler could have been compressed at the input in order to avoid overloading the cheap ADCs and raise small signal above their (high) noise floor. If that's the case, you're less likely to see issues with 16 bit output because the signal is already butchered.

This is why 16 bit places fundamental constraints on the design of instrument amplifiers/modelers. The dynamic range is simply not there for clean tones.

zEr0 · #8 04-15-2008, 10:55 AM

I think Scali realizes he's not actually improving the quality of the Zoom's audio signal but his driver allows him to use it at the same with his other higher quality audio devices.

I'm no expert by no means but it seems to me like a good idea.

microdmitry · #9 04-15-2008, 11:50 AM

I'm not arguing with that. I'm just saying that Zoom's output, if it's 16 bit quantized, is already screwed up beyond repair.

Scali · #10 04-15-2008, 12:16 PM

Well, that was a lot less in-depth than I was hoping for.
I think you also overestimate the dynamics of a guitar. Generally guitarists will use a tube amplifier or a solid-state device to decrease the dynamic range on clean tones, because otherwise it's hard to keep the guitar present in the total band mix (and look, Zoom was smart enough to stick some tubes in there for exactly that purpose!). So the effective dynamic range of a clean guitar is actually quite small. The dynamics of the guitar are not so much in the volume, but rather in the 'colouring' of the sound by varying your picking (did you ever analyse actual CD recordings? Most of them use only a fraction of the total dynamic range that 16-bit offers anyway, even when it's nothing but a clean guitar).
So nice try, but you didn't exactly convince me. You throw some theoretical terms around, but you failed to do any practical analysis.

Aside from that, the Zoom is a modeler which processes at 96 KHz and 32-bit floating point internally. You only get the resulting signal at 16-bit 48 KHz. This is a 'canned' signal, with all the compression and effects you want already applied. Just like how CDs are downsampled to 16-bit 44.1 KHz, this should not be that much of a problem, assuming you're doing the pre-processing (and downsampling) correctly. So there is nothing to 'repair' anyway.

zEr0 gets the right idea... It's just a tool that allows you to easily integrate this device with other high-quality media... Media that probably needs 24/96 more than a preprocessed guitar signal needs.
The 'canned' signal of the Zoom G9.2tt already sounds good anyway, and doesn't need to be 'repaired' in the first place. I don't know where you got that idea.

microdmitry · #11 04-15-2008, 01:18 PM

The reason why dynamic range of CDs is "quite small" is because people demand evenly loud music these days. So sound engineers compress the heck out of everything.

>> The dynamics of the guitar are not so much in the volume,

That too is a lot less in-depth that I was hoping for. Did you actually measure the dynamics of, say, a chord strummed really hard, compared to barely touching the string? Because that's the kind of dynamic range you'll ideally want to handle.

>> So nice try, but you didn't exactly convince me.

Convince you of what? That Zoom is stupid for having a 16 bit output? That it doesn't sound any good because of ****ty ADC/DAC/DSP and compression required to make it all work somewhat passably? I thought that's already well established. This is why modelers (except AxeFX) don't behave properly when you reduce the volume of your instrument using the volume knob or pedal. They don't "clean up" the way tube amps do.

I don't question the utility of your ASIO driver, god forbid. I'm questioning the technical merits of your Zoom modeler. You can't un-butcher the signal. This simple point seems to escape you somehow.

Scali · #12 04-15-2008, 01:32 PM

Quote: