Meridian Unplugged - Powered by MrTechGuy.co.uk

And here is #3 (with aliasing) compared with #5 (no aliasing):

sox #3:



sox #5:



And Meridian's:

Great stuff VK, it would be so helpful, when your done and settled, to give us a crash course on Filters for Dummies becuase most of this is flying by my brain.

Even though ears make the final say I'd like to understand it better technically, but I'm pretty lost at the moment. Partly because I busy on other things at the moment.

Thanks. As you know, I like to understand this stuff technically, before having a listen, usually as means of maximizing my hifi ��s or $$s.

I'm pleased with today's progress: achieving Ayre/Meridian levels of ringing without introducing aliasing is a step forward, though I have nagging concerns from information published by Ayre and Charles Hanson (on this forum) that they achieve their results with more bandwidth. Having said that, the frequency response curve that I posted suggests that the sox bandwidth parameter may not translate directly into the numbers that I calculated (which is a good thing).

TBH, I could stop here. However, I would like to have a listen to an equivalent of Ayre's MP "Listen" filter simply because it is regarded as a state-of-the-art minimum phase filter and because it isn't Meridian.

For years, for us Meridian afficionados, the only way of getting Meridian levels of performance was by using Meridian sources. I believe that the Transporter/SB/sox combination has liberated us, if we choose to go down that route. The Transporter isn't cheap in absolute terms, but it is in relative terms and is a very very good piece of kit which gives me a tremendous amount of flexibility. Whether or not I'm right or wrong, I'm currently viewing the Transporter/SB/sox as a programmable 808.3 - a low jitter device with whatever DSP I choose. Doing my own minimum phase upsampling with a very low jitter transport allows me to compare digital processing algorithms that I would otherwise never be able to compare. (Oddly enough, I can't afford to buy both of Meridian's and Ayre's top level offerings...)

You could post a "calling Charles Hansen" on Audio Asylum Digital Audio forum and ask him to look at this thread. (Email of course doesn't work with him.)

In relation to earlier comments about Ayre, I too don't remember seeing any negativity here, just a deafening silence when I've mentioned them.

Personally I remain rather interested in what they do, particularly in the QB-9 for my Stax headphones setup which sits on the desk in my office/practice room. It just seems rather expensive in comparison with a putative upsampling/apodizing 2011 Ensemble (which could then feed my Staxes through the a analogue zone output, and the main system in the next room via SPDIF). I'm hoping the Ensemble I dream of comes into existence before I have the money for the QB-9...

Sorry, I've taken you off topic, but to go back on, try to get Hansen's attention. He seems a willing sparring partner in such discussions.

I could, of course, PM him on AVS. I'm sure he would be interested given his previous posts here.

Hello -

Virus Killer let me know about the interesting project you are working on here. I must say that this is a very nice forum with all sorts of interesting topics. I wish I had more time to visit and learn about the things you folks are working on.

I briefly skimmed some of the posts concerning creating a minimum-phase filter with the computer program Sox. For those who really want to understand this more, I would suggest downloading a free program used by Ayre (and others) to design digital filters. I think that it is far more powerful than Sox and it also has a far more intuitive user interface. This will allow you to understand the effects of changing the stop-band rolloff, the sharpness of the corner, and things like that:

http://www.mds.com/download/qed2000demo.zip

This package will do everything except spit out the actual coefficients themselves. For that privilege, you need to pay $2000. But now you can play with different filter parameters and not have to guess what they are. Eventually you will have to translate them to Sox or some other PC-operated program.

Hopefully this forum will notify me of any replies, as otherwise I will forget to check back...

Cheers,
Charles

Thank you, Mr. Hansen.

p.s.
It probably won't spit the coefficients, but I wonder what the output will be when one feeds it a single unit impulse input signal ...

It will draw you a pretty picture of what the impulse response will be. This is extremely useful.

Let's say that you create a target response in the MDS (Momentum Data Systems) program. Even if you had the full version that would spit out the coefficients, it may be that something like Sox won't even let you input coefficients. But you can fiddle with the parameters in Sox until you get a similar looking impulse response.

And the impulse response describes a digital filter completely. So you can use MDS to play around with different parameters far more easily than with Sox. Then use Sox (or some similar program, they must exist) to create the desired filter function.

Hi Charles,

Thanks for your input. I've taken a look at the MDS program, but it's highlighted something that I've known for a while that I don't understand. Specifically, the relationship between a FIR filter and a change in the sample rate.

Does the FIR filter work in conjunction with the sample interpolation or are the two combined? In MDS, I have to set the FIR sample rate to 88.2k in order to try out filters which have very slow roll-off down to 44.1k. There's nothing where I can say "I want to go from an input of 44.1k to an output of 88.2k".

This is relevant because, although sox *does* have a FIR engine into which one can feed an arbitrary number of filter coefficients, I can see no way to change sample rate without invoking the resampling routine itself, filter and all.

Also, as there appear to be many different ways to perform the re-sampling interpolation, this would suggest that the interpolation *might* be a separate process.

Joel

It's possible that some filter design applications will not let you go into non-Nyquist (aliasing) territory, hence the requirement for a higher Fs if you want to roll-off slowly. It's just a supposition, though.

On the subject of resampling and filtering being separate processes, it's worth noting that one can arrange things so that the maths operations involved are identical, and given the same filter order for both, these can be compounded an performed at once.

When using MDS, just set the sampling frequency to the desired oversampling frequency. It gives you about eight graphs, only two of which are useful -- frequency response and impulse response. Just close the other ones and then select "Tile" and it will show just those two graphs.


I don't know if you can do this with Sox, but the standard way to perform oversampling is with "zero stuffing". So with 2x oversampling, just insert a sample with a value of zero in between each existing audio sample.

Obviously this technique only works for integer values of oversampling (2x, 4x, 8x, 16x, etc). Then run the data through the low-pass digital filter that you've created in MDS. Finally, you will have to shift the output bits right or left so that they have the right amplitude. For example, if you are doing 2x oversampling, you have decreased the average signal level by half (or -6 dB). Then you would have to shift all the bits one bit toward the MSB side to restore the proper level.

I don't know if Sox can do all of that. But MDS is unbelievably good for getting a feel for the tradeoff between sharpness of the rolloff curve and the amount of ringing. With the latest version of 2000, you can choose to make minimum-phase filters, so it is easy to compare what happens when you just change that one parameter and leave everything else alone.

Hope this helps!

MDS will not let you alias. It gives you a warning and tells you to pick a different parameter.


Implementing the filters is a completely different story. There are a zillion ways to do it. In hardware you can use a DSP chip or an FPGA, or even a microprocessor if the algorithm is a simple one. In software there must be even more ways. We've always done it in hardware, as we don't want to put a PC inside our products. (However, some people do just that. I believe that one of the Naim products is basically just a PC and I'm sure there are others. The Olive music servers are very stripped down PC's running Linux.) And then there are zillions of different algorithms. As a trivial example, when oversampling by zero stuffing, there is no point to calculate the result of the product of the coefficient and the zero-valued sample. You already know in advance that the result will be zero.

But I think that is a thread for a different forum. I just wanted to point out a tool that is probably more flexible than the pre-determined choices in Sox that might help get the most out of Sox.