|
Audio Asylum Thread Printer Get a view of an entire thread on one page |
For Sale Ads |
24.229.63.32
I just got a Cal Labs Alpha DAC. WOWIE FREKIN WOW WOW. If I said I was hearing a whole new collection of REAL music, that would be an understatement. I'm using an $8.00 sony dvd player as a transport and it still sounds phenomenal. Try TUBES they just make music.
Follow Ups:
Hi Charles, Todd and others, I'm an audiophile without any technical background. Recently I came across this discussion that I have read with great interest. For me a CD player or DAC should only do one thing: getting the best sound from CD that is closest to reality. If I understand the explanation correctly, DAC chips don't have the biggest impact on sound quality but the filters that are being used to cancel the high frequency noise artifacts? Well, I wonder if it is possible to design a filter that can get rid of those noise artifacts without damaging the wave shape and/or introducing gross phase shifts or is this something that is technically impossible within the boundaries of Red Book?
Btw, I don't care if a filter rolls off frequency with 2dB at 20kHz. I only care for good and realistic sound.
Chris Hie (Netherlands)
dazzdax@xs4all.nl
"If I understand the explanation correctly, DAC chips don't have the biggest impact on sound quality but the filters that are being used to cancel the high frequency noise artifacts?"
The chips do have the biggest impact on sound quality, in my opinion, because the filters are already in the chips (unless an upsampling chip is used prior to the DAC chip, in which the upsampling chip's filter takes over), and are not exactly optimized for the best sound quality possible. (Although the new "minimum-phase" filters recently introduced is a giant step in the right direction, in my humble opinion.)
The filters themselves block, to varying degrees, the ultrasonic reflections, and push out the first alias much farther out into the ultrasonic band. They're used mainly because they're effective at blocking out the initial alias components (prior to any filtering) in the ultrasonic band but very close to the audio band. (These reflected signals can be called "noise", since they're not musically related to the desired part of the signal. Although the noise is correlated to the signal.)
"Well, I wonder if it is possible to design a filter that can get rid of those noise artifacts without damaging the wave shape and/or introducing gross phase shifts or is this something that is technically impossible within the boundaries of Red Book?"
That, my friend, is the confounding issue. Since half the sample rate (aka "Nyquist frequency") is 22,050 Hz, there is very little room to execute an effective filter to preserve bandwidth to 20,000 Hz. And the best filters in my opinion don't necessarily block everything at 22,050 Hz and above most-effectively, but preserve/reconstruct most of the waveform fidelity of what the signal was prior to digitization.
I never could have imagined
Now I can never listen to music any other way, in fact I dont see how or why any DAC manufacturer would even want to make anything 'other' than a NOS DAC.
Karma Means Never Having To Say You're Sorry
> > I dont see how or why any DAC manufacturer would even want to make anything 'other' than a NOS DAC. < <
After years of *planning* to get around to it, we finally did some listening to NOS (Non=OverSampling) DACs this past fall. While they had their charms, we found that there were better solutions overall by using a digital filter (depending, of course, on the particular digital filter). See link below.
But as long as your system is making music that sounds great to you, then I'd quit while I was ahead.
Frequency and impulse response plots provide little information regarding a DAC’s performance with music. Frequency response uses steady-state sine waves and very few DACs have trouble with that kind of signal. But, if you add an amplitude envelope the landscape changes.
Here is a tone with a linear up/down ramp processed by DAC Y (non-oversampled.)
Here is the same signal processed by DAC A (8x oversampled, linear phase FIR, delta-sigma modulation.)
The ringing is evident even though the signal slowly ramps up/down from/to zero. Worse yet is the dynamic compression. This is common to all FIR filters. FIR filters essentially replace each sample point with a weighted average of the surrounding samples.
An impulse is noise and a DAC’s response to noise is not very informative. However, if you add a frequency component to the impulse all kinds of unpleasant things show up.
Here is a two-cycle sine wave burst processed by DAC Y.
Here is the same signal processed by DAC A.
Here is the same signal processed by DAC W (8x oversampled, minimum phase FIR, delta-sigma modulation.)
Unlike the natural decay of a note, the ringing is enharmonic. What’s more, the reproduced signal is also off-pitch.
In all cases the signal frequency is at least seven octaves below the Nyquist frequency.
Wow, what an interesting post. What are signal frequencies or X scales of you pictures? Gonna have to do some playing with this!
Regards, Rick
Think I will stay with my NOS DAC...sounds great with music. Don't listen to sine-waves much.
Karma Means Never Having To Say You're Sorry
Get yourself a good text book on Digital Signal Processing and study it. After you have done so, then you may have a good idea of what is wrong with your post. A reference is listed below, but you may need to learn some math before you can understand it.
I will point out one thing. If the signal were 7 octaves below the Nyquist frequency, then there would be 128 cycles of ringing for each cycle of your "signal". One cycle of your signal occupies about 50 horizontal pixels in your photos. So any ringing from a filter that cuts off near the Nyquist frequency would not be visible in your pictures.
Tony Lauck
"Diversity is the law of nature; no two entities in this universe are uniform." - P.R. Sarkar
- http://www.amazon.com/Discrete-Time-Signal-Processing-2nd-Prentice-Hall/dp/0137549202/ref=sr_1_1?ie=UTF8&s=books&qid=1234322188&sr=1-1 (Open in New Window)
Mr. Lauck,
You’re the one who needs tutoring in the fundamentals of digital audio. You should learn that the Nyquist frequency is one half the sample-rate and the sample-rate is defined as the rate at which the samples change. The Nyquist frequency is important because it defines the highest frequency that can be digitally recorded and it also defines the cutoff frequency of the ideal reconstruction filter. The latter is the sole justification for up/oversampling because it raises the sample-rate, and hence the Nyquist frequency. That relaxes the requirements of the reconstruction filter.
When I said the test signal was at least seven octaves below the Nyquist frequency I was being absolutely precise and I assumed the superstars of digital audio, like you and Todd Krieger, would know exactly what I meant. I assumed too much.
If music is digitally recorded at a high sample-rate, say 192K, the Nyquist frequency is 98KHz and, prior to digitization, the signal must filtered to remove all frequency components above 98KHz. If the music is subsequently down-sampled to 44.1K for CD production the new Nyquist frequency will be 22,050Hz and, prior to down-sampling, the signal must be filtered to remove all frequency components above 22,050Hz. That’s usually done with a digital FIR filter. Now, when the CD is played back through a NOS DAC, the Nyquist frequency remains at 22,050Hz and the reconstruction filter, either explicit (RLC network) or default (amp, speakers, and ears), must remove all audible frequency components above 22,050Hz. However, if the CD is played back through a digital oversampling filter and/or a delta-sigma DAC, the Nyquist frequency is raised by the final oversampling ratio of the DAC/filter.
The DAC A and DAC W both have specified output sample-rates of 128 times the input sample-rate. Hence, the Nyquist frequency of those DACs is seven octaves above the Nyquist frequency of the 44.1K input sample-rate, or about 2.8MHz. 20KHz is generally accepted as the upper limit of the audible frequency range and 20KHz is at least seven octaves below 2.8MHz.
Incidentally, both D-S DACs can reproduce a perfect, steady state 20KHz sine wave from a 44.1K sample-rate input. They only have problems dealing with discontinuities in frequency and/or amplitude. That is the nature of FIR filters where each new sample is computed from a weighted average of the surrounding samples. The filter design determines the number of surrounding samples included in the computation and the weighting applied to each included sample.
Here is a one-octave trill processed by DAC Y.
Here is the same signal processed by DAC W. It’s not pretty.
Here the test signal is at least eight octaves below the Nyquist frequency of the respective DAC.
"When I said the test signal was at least seven octaves below the Nyquist frequency I was being absolutely precise and I assumed the superstars of digital audio, like you and Todd Krieger, would know exactly what I meant. I assumed too much."
When a DAC converts discrete samples into a continuous waveform a filter is needed to eliminate the images that appear above the Nyquist rate of the input sample rate. Whether this filter is done in the analog domain or by digitally upsampling to a much higher frequency, the same filter properties are required for reconstruction of the output waveform. The stop band of the filter must start at the Nyquist frequency corresponding to the input sample rate. Ringing in this filter will be determined by the filter response in the transition band located between the pass band and the stop band. In other words, ringing will be at or near the input Nyquist rate. The output sample rate is irrelevant to this discussion, although it does affect the filter design.
Even if a filter "rings" it will not produce ringing in the output unless it is provoked with a signal near the ringing frequency. So when you report that the two cycle waveform results in ringing at a non-harmonic frequency, this is not correct. What is happening is that the input waveform has energy at all frequencies up to the Nyquist frequency of the input sampling rate. Two cycles of a sine wave can be modeled as the product of a continuous sine wave and a pulse waveform (which is the sum of a positive and negative step function). The pulse waveform has energy at all frequencies and hence the product has energy at all frequencies. It makes no sense to say that this ringing is non-harmonic. The filter is not adding any frequency components that were not in the original signal.
"Here is the same signal processed by DAC W. It’s not pretty."
Without seeing the input, it is impossible to tell if DAC is doing a proper job of creating output. By input, I mean the actual samples, that is a bunch of dots on a graph, not a bunch of lines someone or something drew that connect the dots. Also, it is hard to tell from low resolution images how much of the ugliness has to do with image processing rather than audio processing.
Tony Lauck
"Diversity is the law of nature; no two entities in this universe are uniform." - P.R. Sarkar
It makes no sense to say that this ringing is non-harmonic. The filter is not adding any frequency components that were not in the original signal.
It makes perfect sense. The fact that the frequency of the ringing is not an integer multiple of the frequency of the signal, by definition, makes it enharmonic. The fact that none of the frequencies output by the filter match the frequency of the input signal, at any time, by definition, makes the output of the filter enharmonic and off pitch. The fact the filter is outputting non-zero samples during periods of time where the input samples are zero, by definition, means the filter is adding frequency components not in the original signal.
The input signal consists of three segments. The first and third are silence and the second consists of two complete cycles of a sine wave. The instantaneous slope at every point of the input signal, plus/minus the quantization error, matches the instantaneous slope at each corresponding point in those idealized segments. The fact the filter sees all frequencies between 0 and Fs/2 is a problem with the filter and not the input signal. The fact that the filter computes each sample point as a weighted average of surrounding samples is a problem with the filter, not the input.
You don’t really need to see the input sample points to conclude the output of DAC W in response to a 10KHz trill is pure fiction.
Here is the same signal processed by DAC A.
Here is the exact same signal, i.e., the exact same frequency, processed by DAC N, a NOS DAC with a minimal, non-brick wall, analog reconstruction filter.
"The fact that the frequency of the ringing is not an integer multiple of the frequency of the signal, by definition, makes it enharmonic. The fact that none of the frequencies output by the filter match the frequency of the input signal, at any time, by definition, makes the output of the filter enharmonic and off pitch. The fact the filter is outputting non-zero samples during periods of time where the input samples are zero, by definition, means the filter is adding frequency components not in the original signal."
What you stated is true for the instantaneous case, but not true for the averaged case. You've basically described the "beating" or "modulation" that takes place with signals close to Fs/2. A long-enough "windowed sinc" filter actually alleviates this problem to a degree. (Provided the signals are continuous.) But it causes other problems.......
The main issue with Redbook CD playback is depending on the choice of filtering (or no filtering), there is a tradeoff between good transient performance (less ringing) with HF modulation (whose averaged amplitude results in the "slow-rolloff" frequency response) and good frequency response (related to reduced HF modulation) with poor transient performance (ringing). The former case is with NOS or limited filtering, the latter case is with classical (brickwall/"sinc") filtering.
"The fact that the frequency of the ringing is not an integer multiple of the frequency of the signal, by definition, makes it enharmonic."
There is no single frequency of this signal. It consists of a range of frequencies. Linear time invariant systems do not output frequencies that are not present in their input. They only change the amplitude and phase of the input frequencies.
I gave you bad advice when I suggested you read a book on Digital Signal Processing. You are not ready for this book. You should take a course on Signals and Systems. Perhaps the following link will be helpful, I haven't checked it out, but at least it is free.
Tony Lauck
"Diversity is the law of nature; no two entities in this universe are uniform." - P.R. Sarkar
Hey Tony, in case you forgot, the goal of the digital signal in question was to recreate two cycles of a 20KHz sine wave. Apparently, all your high-powered digital filters can’t do the job. How many more cycles does it take? Three or four…hundred….thousand….million? Regardless, what ever the number there will always be a discontinuity at the beginning and the end of the sample stream at which time the FIR filter will output undesired tones that are enharmonic to the frequency you are trying to reproduce. A NOS DAC can’t quite do the job either but it comes a lot closer.
Here is the very same two cycle, 20KHz signal processed by DAC N. It’s got about the right number of cycles and, most importantly, the right frequency.
Here is the same signal processed by DAC L, an unfiltered NOS DAC. It comes the closest to the truth provided your amp, speakers, and ears can filter out all the undesirable but audible frequency components of the stair steps and all the audible intermodulation distortion that may be produced by the inaudible frequency components of the steps.
Most audiophiles proudly claim to hear things that can’t be measured or observed with an oscilloscope. On the other hand, you claim the digital filter anomalies that can be observed with an oscilloscope are, nonetheless, inaudible. How convenient. You may not hear them but others do. In addition to the anomalies I documented, an FIR filter’s propensity to average adjacent samples sucks the life out of the music. Natural acoustic music has an unmistakable texture and that texture is reduced each time the music passes through a FIR filter. With the ubiquity FIR filters in all stages of music recording and production, most modern digital recordings sound more like the lifeless output of a digital synthesizer than real music created by humans singing and playing acoustic instruments.
"Hey Tony, in case you forgot, the goal of the digital signal in question was to recreate two cycles of a 20KHz sine wave. Apparently, all your high-powered digital filters can’t do the job."
It is not possible to represent 2 cycles of a 20 kHz signal in the 44.1 kHz format. Regardless of the DAC, you can't do it. If you don't like it, then blame Sony / Philips. This issue would be very clear if you were to try two cycles of a 200 Hz signal in the 44.1 format. All of the DACs would look pretty good.
"It comes the closest to the truth..."
You are expressing a preference in a multi-dimensional space. Neither DAC provides a very good approximation to your idealized waveform, and you have given no basis other than your eyes for choosing between the two alternatives. Others will look (and listen) and reach a different conclusion. In addition, with a different waveform the preferences might be different.
"Regardless, what ever the number there will always be a discontinuity at the beginning and the end of the sample stream"
That is true, but it has something to do with the nature of digital audio and not the implementation of a playback filter, whether done in the digital, analog or biological domains. The Sampling Theorem applies only to band limited signals. There are no non-zero band limited signals of finite duration. So according to your criteria, no digital audio system can perfectly reproduce your waveform. I am not going to accuse you of being a troll, but if you really believe that digital audio is inherently no good, then you have no business being in this forum.
Your two cycle test signal exists only in the platonic world of numbers. It does not exist in physical reality. There is no way to create sound waves that start and stop instantaneously and no way to transform those sound waves into electrical signals that start and stop instantaneously. If you do not believe this, then my advice would be to take a course in Physics.
Tony Lauck
"Diversity is the law of nature; no two entities in this universe are uniform." - P.R. Sarkar
> > Hey Tony, in case you forgot, the goal of the digital signal in question was to recreate two cycles of a 20KHz sine wave
That is NOT a band-limited signal, and can NOT be correctly recreated by a system sampling at 44.1. Not NOS, not upsampled. Never. Until and unless you understand the very important implications of that, further discussion is completely meaningless. Bickering about filters etc is completely meaningless if you violate the basic primary assumption of sampling theory. THAT is why Tony correctly pointed you at signal theory references.
This is probably the millionth time someone has advanced this argument on this forum. It was wrong the first time. It's wrong now. It'll be wrong next month when someone else tries it again.
I didn’t get it when digital audio was introduced as Perfect Sound, Forever because Sampling Theory told us so and laboratory measurements with steady state test tones confirmed it. The obvious audible limitations of 16 bits @ 44.1K were dismissed as theoretically inaudible . And so it goes. The wizards have further perfected Perfect Sound, Forever with digital filters and other special effects and, again, we are told the obvious audible imperfections are theoretically inaudible and the wholesale modification and averaging of recorded samples is described as statistically insignificant .
When I post plots showing what I consider significant anomalies in FIR filters I am branded as an ignoramus who doesn’t get it because the signals I created were not band-limited. Yet, an audio component manufacturer is praised after posting plots showing his FIR filter’s impulse response. As far as I know, impulses are, by definition, not band-limited.
OK, now I get it. Non band-limited test signals are good as long as they further the cause but any non band-limited signal used to criticize the cause is deemed invalid. As Tony just pointed out, becasue of my views, I have no business posting in this forum. I can take a hint.
I share your dislike of 44.1/16 audio and the despicable marketing campaign associated with it. I would be delighted if Sony and Philips were to go out of business (assuming their non-management employees found suitable alternative employment).
Unfortunately, you undercut your own case by using pseudo-technical arguments that do not conform to known theorems of mathematics, established physical laws, and common engineering principles. I have offered suggestions as to how you could increase your knowledge. If you do not have the time, interest, or talent to do this (and we are talking about hundreds or even thousands of hours of study) that's fine, too. But you might find it more productive to confine your remarks along lines where you do have expertise. I have no way of knowing what these might be, but my experience has been that nearly everyone has some unique talent and expertise.
Tony Lauck
"Diversity is the law of nature; no two entities in this universe are uniform." - P.R. Sarkar
There does seem to be a problem with the band limiting of the images shown in the original post, hard to tell what the sample rate is though. And I agree, there also seems to be some confusion on the Nyquist frequency that applies to the digital filter, as compared to the output sample rate Nyquist, which is usually taken care of with just a simple analog post filter.
Regardless, this is on a little different subject, and while I know there are some here that will disparage any type of upsampling applied to the signal before the standard oversampling digital filter, if the upsampling ratio is high enough, say to 4x to 5x, aren't you able to reduce the filter slope quite a lot since the upsampler will have a stop band that can stretch from around 20K all the way up to the new sample rate Nyquist, instead of just up to the original Nyquist, and the digital filter could be set to act at the new sample rate, with whatever filter type it is set for?
"There does seem to be a problem with the band limiting of the images shown in the original post, hard to tell what the sample rate is though."
Looking at the pics, the ringing frequency is obviously close to the signal frequency, which suggests the signal frequency is close to half the sample-rate frequency, aka "Nyquist" frequency. It is not "seven octaves below" as claimed.
"And I agree, there also seems to be some confusion on the Nyquist frequency that applies to the digital filter, as compared to the output sample rate Nyquist, which is usually taken care of with just a simple analog post filter."
You have this correct- The digital filter acts on the input sample rate, the analog post filter acts on the digital filter's output sample rate.
"Regardless, this is on a little different subject, and while I know there are some here that will disparage any type of upsampling applied to the signal before the standard oversampling digital filter, if the upsampling ratio is high enough, say to 4x to 5x, aren't you able to reduce the filter slope quite a lot since the upsampler will have a stop band that can stretch from around 20K all the way up to the new sample rate Nyquist, instead of just up to the original Nyquist, and the digital filter could be set to act at the new sample rate, with whatever filter type it is set for?"
Yes, provided the oversampling DAC still oversamples at the higher rate. (Not all do, but the Wolfson apparently does.) The only advantage is it would require an even smaller post filter, but I think a second oversampling stage is totally unnecessary, since it would require intensive number-crunching for a filtering function that's non-critical. It's only critical the first time, since the Nyquist frequency is initially so close to the top of the audible 20-20k band.
My thinking is that a good first stage filter can be followed by a mediocre second stage filter, which will be much less critical because it is operating way out of the audible band.
Whether a two stage approach or a one stage approach sounds better will depend on the particular filters involved. Note that the first stage filter increases the bit depth of the signal going to the second stage filter, so upsampling CD material to higher rates is probably a poor idea if you are using a 16 bit DAC.
Another advantage of the two stage approach is that the critical first filter can be implemented by software, making it much easier to tweak.
Tony Lauck
"Diversity is the law of nature; no two entities in this universe are uniform." - P.R. Sarkar
I've seen conflicting information on what the filter requirements are when doing sample rate conversion to a higher rate. Some say that you need to apply the anti-alias filter for the original Nyquist frequency, while I think the more correct answer is that the slope of the filter can now be reduced since the stop band is wider, stretching to the new sample rate Nyquist frequency. Is this the correct answer? If so, that seems like a big advantage, especially if you skip any additional "standard" digital filtering.
The idea behind digital playback is that if you sample a *low-pass filtered* signal, then that signal can be reconstructed *exactly*. The low-pass filter on the record side is called and "anti-alias" filter.
To do this means that the playback system *also* has to have a low-pass filter. This filter is called a "reconstruction" filter. If everything is done right, then the playback signal will *exactly* duplicate the original *low-pass* filtered signal.
So there are a couple of gremlins hiding here.
The first is that adding a low-pass filter to an audio signal will change its sound. Period.
The lower the frequency of the filter, the more it changes the sound. That is why CD sucks compared to vinyl, which has no explicit filters in the signal path. That is also why the "high-res" formats had a chance to sound better than CD.
The second gremlin is that since CD is so restricted in how good it can ultimately sound that there are a lot of people that take liberties with the playback "reconstruction" low-pass filter.
But digital oversampling filters have NO ADVANTAGE over an analog filter except that you can make a very complex one much more inexpensively than an analog one. They are simply trading a complex digital filter for a complex analog filter.
THERE IS NO FREE LUNCH!
The filter in a DAC is used to remove images created when the samples are converted into a higher sampling rate domain or to analog (think of analog as an "infinite" sampling rate). The images exist because of the original samples—they are not there because of the higher sampling rate. A low pass filter is needed to remove them (or at least to reduce them).
The slope of the filter depends on the difference between the top of the pass band and the bottom of the stop band. The top of the pass band depends upon the desired frequency response. The bottom of the stop band must be at or below the Nyquist frequency for the sampling rate of the input signal. The output sampling rate is not relevant; there might not even be an output sampling rate if the filter is an analog one. The stop band for CD playback starts at 22,050 Hz and extends to infinity.
Tony Lauck
"Diversity is the law of nature; no two entities in this universe are uniform." - P.R. Sarkar
:-)
I had always (mistakenly) assumed that upsampling from say 44.1K to 192K didn't really have any big advantage over standard oversampling, but Ted Smith pointed out that it would get rid of the need for a brickwall filter. See below thread ...
Ted is usually spot-on with his posts. But the one you linked was wrong. The first stage of the digital filter *has* to have a brickwall filter (or something very close to that).
Interesting photos.
At this point, I'm interested and also skeptical.
I'd like to know how those waveforms were generated and captured. You should also be kind enough to share the specific DACs that were measured. The reason is that I would like to reproduce your experiments. Your last graph claims that the DAC uses an 8x minimum-phase FIR. I am not aware of any product on the market with that combination of features.
We don't own an arbitrary function generator. But that shouldn't be a problem, as a computer program could generate arbitrary waveforms and then be used to burn CD's as test discs.
If you have the WAV files, please e-mail them to me and I will burn a disc and we can use it to test the new MP players.
The thing to remember is that all FIR filters are not alike. So I don't think it's a fair statement to say that they are *all* fatally flawed.
The other thing to remember is that we chose the filters based on the result of listening tests. Those tests included listening to non-oversampling (filterless) designs.
I don’t have an arbitrary function generator either. Who needs one for digital audio? All you need is a computer. If you have computer, you can generate wave files and burn CDs as easily as I can.
DAC A is an AKM and DAC W is a Wolfson. I used your term “minimum phase” because, as you depicted on page 3 of your white paper, the Wolfson DAC minimizes pre-ringing at the expense of post-ringing.
All FIR filters are fatally flawed because they compute the value of each sample from a weighted average of the surrounding samples. That works perfectly for steady state sine waves but not music. Music is anything but steady state. Even the simplest melody starts, changes from note to note, and than stops. More often than not, the changes from one note to the next are abrupt. There are also continuous changes in pitch and amplitude in the fundamental and harmonics of each note as it is sounded. As I see it, FIR filters strive to create the smoothest sine wave output using the input signal as a guide. In the end, there is no direct correlation between the input and the output: Just a reinterpretation.
See also my reply to Mr. Lauck, below.
I am interested in replicating your work.
My e-mail account has no limit on the inbox size. So simply e-mail your test signals to me.
I would also like more information on your test setup. It sounds like you made your own D/A converter that could accept different DAC chips. Is that correct? Or were you using the manufacturer's eval boards?
There are a lot more questions, but let's just start there. Then I can run your test signals through our DAC and see what results we get.
Thanks.
I used your term “minimum phase” because, as you depicted on page 3 of your white paper, the Wolfson DAC minimizes pre-ringing at the expense of post-ringing.
Minimum phase is a well known term to those skilled in the art of processing signals. That you would attribute this term to Mr. Hansen shows your ignorance of this field.
Tony Lauck
"Diversity is the law of nature; no two entities in this universe are uniform." - P.R. Sarkar
"In all cases the signal frequency is at least seven octaves below the Nyquist frequency."
Based on the pics, I'd say the signal frequency is *very* close to the Nyquist frequency. The ringing components are roughly the same frequency of the signal components. If that ramp up/ramp down waveform were seven octaves below the Nyquist frequency, the ringing components would be roughly seven times the signal frequency. But the ringing components appear to be roughly the same frequency.
But on the other hand, for so long of a time, this problem has been an elephant in the room that the audio engineering community had chosen to ignore.
Charles:
Your MP white paper has to be one of the best ever at explaining digital filtering so that just about anyone could understand it. The text was kept simple and the graphs clearly supported it. Two questions:
1. How come no one thought of this before? As stated in the text, part of your idea was based on Peter Craven’s AES paper that came out in 2004.
2. How do you think your competitors would respond to this? For example, do you think they would say this is an oversimplification?
Thank you.
Gerry
If you turn back the clock 20 years, Wadia and Theta were the undisputed leaders in digital audio, largely because they used custom filters implemented in Motorola DSP chips.
This started to crumble when the Levinson 30/31 came out. It was widely received as the best sounding digital product, yet used an off-the-shelf filter from NPC (later changed to the PMD-100 HDCD decoding filter).
With the rise of one-box players, people largely dropped the custom DSP-based filters. Now we are resurrecting them, as John Swenson notes in his post, with FPGAs. The learning curve is pretty steep with FPGAs. But we have been using them for years because that was the only way to do the things we needed to do with the video signals in our DVD players.
As far as "minimum-phase" filters, I think some people have played with them, but not in a big way for CD players. For example, there was a digital equalizer from Z-Systems that had minimum-phase filters. But most people implement minimum phase with an IIR. We found that an FIR sounds better, presumably because of the cumulative error as the data makes multiple passes through the filter of an IIR.
Using an FIR for a minimum-phase filter is so unusual that I couldn't find any software or equations that would let us do the de-emphasis with an FIR. So I did it the old-fashioned way -- with trial-and-error, converging on the target function.
And it doesn't matter what the competitors say. What matters is what the customers say. It's a nice story, but it only matters if it actually makes a sonic difference.
"1. How come no one thought of this before?"
The vast majority of D/A processors use an off-the-shelf DAC chip, and there has been no way to get around the built-in digital filter on the chips. (Later chips offered a "slow-rolloff" option, but the Ayre MP is well beyond that.) One must expend a lot of resources to get a custom filter design to market. It takes a full blown company (like Ayre) to pull it off..... There isn't such backing with most low-budget audio companies.
"2. How do you think your competitors would respond to this?"
Companies like Wadia and Krell might provide something similar..... (MSB Tech. might also have a filter like this as well.) But the low-budget companies might be sunk if a chip manufacturer like Wolfson or Burr Brown doesn't come forth with a DAC chip with such a filter. The only alternative would be to implement the filter in software, as a Foobar or Winamp plug-in...... (Talk about the rug being pulled out from under the Sabre DAC.)
Years ago anybody trying to implement their own digital filter either had to design a custom chip (very expensive) or write software for DSP (digital signal processor), less expensive but inefficient.
With the advent of modern FPGAs its much easier to do your own custom filter. An FIR filter in an FPGA is not that complicated. The fun part is in figuring out exactly which filter function to use.
The problem is that we don't really know exactly what to optimize for. There is no equation that will design a filter that "sounds best". What happens is you have a hypothethis about what makes it bad or good, you design a filter that implements that (or DOESN'T implement it), then you listen. Repeat many times.
And to complicate things different people cannot agree on what constitutes the best tradeoff. I've been doing this trying different filters bit for several years now and have had several occasions where I had a group of people listening to the results. I would play a certain filter that I happened to like, and several of the listeners would agree that this was an improvement. But invariably one person whould respond with "thats awfull, what did you do to the sound?" So what do you do? Come up with the filter YOU like the best, or use the one with the broadest appeal of your test audience, or ...?
Then there is the issue of do you use the same filter for all sample rates? The big tradeoffs are for 44.1 since it is so close to the audio band. Does it make sense to use the same filter at 96 or 192? maybe not. I don't think there has been a lot of exploration of this aspect.
John S.
"So what do you do? Come up with the filter YOU like the best, or use the one with the broadest appeal of your test audience, or ...?"
Surely, it must be possible to program a couple of different FPGA filters that one can switch back and forth to one's tastes?
Many other options with low, medium and high rate filter responses.
I haven't paid much attention to the Wolfson parts, because like many other DAC chip makers, they only have voltage output parts. This means that the current-to-voltage conversion is being handled internally. We prefer to do it ourselves outside the chip for best sound quality.
But it looks like the 8741 and 8742 have a lot of filter options. Unfortunately, the data sheets don't give enough information to really tell what is happening with the filters. As usual, they appear to be a cascade of three 2x sections to create an 8x filter. There is a bit more information in the linked AES preprint, but not enough to tell what they are really offering. However, it seems to be a step in the right direction.
... about different filters for different sample rates. I was just pointing out that some of the popular DACs like the Wolfson do have different filters for different rates, in this case there are 3 bands with 5 filter selections for each (though only 3 filters are available for each in hardware control mode). Not meant to imply it's a replacement for a well implemented approach such as yours, but a sign that for the last few years the digital audio world maybe hasn't been quite as dire as some believe :)
Like you say, in the 2005 AES paper they do talk some about the available filter options and some of the reasoning behind the choices ...
AES 119th Convention, New York, New York, 2005 October 7–10, Page 4 of 10
2.7. Interpolation filters for 44.1kHz and 48kHz audio
Since there is no definitive indication that minimum
phase filters sound better than linear phase filters,
particularly given the phase distortion over the audio
band it was decided to implement a series of filters and
allow the end user to decide which filters to use. The
reduced latency of the minimum phase filters may be
ideal for studio situations where a low latency is needed.
Conversely some users may prefer the standard linear
phase filters.
The filters we chose to implement are:
1. linear phase
2. minimum phase with tailored response
3. minimum phase
4. no-alias linear phase
5. no-alias minimum phase
3.1.1. Interpolation filters for high sample rate audio
For the higher sample rates we settled on 5 filters
1. linear phase ‘soft knee filter’
2. minimum phase ‘soft knee filter’
3. linear phase brickwall filter
4. minimum phase apodising filter
5. linear phase apodising filter
Like I said, it seems to be a step in the right direction. I don't know how long those parts have been available, but it's not surprising that they are -- it seems that when the time is ripe for an idea that there is often a lot of parallel development from several sources.
Assuming those parts have been available for a while, I'm surprised that not more people are using them. The only other commercial product I know of that has a minimum-phase filter is the Meridian 808 Mk.II Signature CD player. It has an "apodizing" filter that is quite similar to the "apodizing" filter that is the "Measure" position of our new MP players. As far as I know, Meridian only offers the one filter choice and no other selections.
When we put the "Listen-Measure" filter selection switch on the back, it was for a reason. I *hate* user adjustments like that. To me, it is an admission from the manufacturer that says, "We couldn't figure out what sounded best, so we'll let you fiddle with it and see if there is something that you'll like." However, one of magazines blasted a Pioneer CD player with its slow roll-off "Legato Link" filter (similar to our old "Listen" filter) for not having ruler flat frequency response. So we a selector switch (but put it on the back!) just so the magazines (especially the German ones!) could measure and not have a heart attack because it was 2 dB down at 20 kHz. But it's called "Listen" for a reason -- that is what sounds best and that's what you should use when you are listening to the danged thing.
All of this nonsense is strictly due to the limitations of the CD format. At higher sample rates, the roll-off from a "slow roll-off" filter is strictly academic. It really doesn't matter if something is a few dB down at 48 kHz or 96 kHz. But that is one thing that has slowed down the pursuit for better sound -- all pro gear is largely sold on the basis of specs. So they all used brickwall filters to make sure that their gear would be dead flat to as high a frequency as possible.
As some of the posts in this thread point out, it's not clear that filters (record or playback) are even necessary at high sample rates. I think that a 192/24 system with no filters could sound absolutely stunning. But as far as I know, it hasn't even been tried. I think the main reason SACD sounds as good as it does is because there are no filters on the record side. The problem is that they have to use fairly steep filters on the playback side to keep the high frequency noise from overloading most electronics.
The cheap little DacMagic that has been talked about all over the boards uses two of the 8740s, and I think Onkyo has incorporated that part in a lot of their products now. Probably more, they are getting pretty popular. Minimum phase designs for the masses :)
But yea, it would be nice if they offered something that had a better DAC stage along with current outputs instead of having to use the switched capacitor/opamp type voltage outputs. Nothing better than a ladder and simple current mirror in my experience for I/V conversion with PCM sources. Or even if they could just offer the digital filter outputs so one could use their own preferred DAC stage, that would be nice too, and would probably sound pretty good hooked to a PCM1704 or something like that.
I think that the 8740 offers a slow roll-off, but not the minimum-phase features. I am unaware of any commercial product that touts minimum-phase filters aside from ours and the recently released Meridian 808 MkII Signature CD player.
The Burr-Brown DF1704 was the first commercially available chip that I know about that used slow roll-off algorithms. We used that in our D-1 DVD player starting in 1999. (Wadia was the first complete DAC, available in 1989 or so.)
Looks like you don't get all of the advanced filter selections except on the 8741/42. It appears the DacMagic is bypassing the digital filter and using their own DSP approach like you to get the minimum phase design. Arcam is supposed to be using the the 8741 with the 15 filters in some of their products, but I don't know if they make the filter options available to the user.
Don't the AKM D/A convertors that lots of people rave about use a minimum phase design as well? I haven't really been following too closely all of the new designs.
" I am unaware of any commercial product that touts minimum-phase filters aside from ours and the recently released Meridian 808 MkII Signature CD player. "
See link.
I had heard about the DAC Magic, but did not realize that they offered a minimum=phase option.
They are using a low-cost DSP chip from TI with algorithms from Anagram of Switzerland. It looks like they are doing ASRC (Asynchronous Sample Rate Conversion) of all sources to 192/24 and then feeding it to the Wolfson DAC chip. I would assume that the Wolfson further oversamples it another 8x, but it is not clear if this is the case or not.
Guys,
I would think they are using the 8x input which bypasses the internal filters of the Wolfson dac.
Thanks
Gordon
J. Gordon Rankin
Charles,
There's also a mini-review of it in the March "Sam's Space" if you haven't seen it yet.
"They are using a low-cost DSP chip from TI with algorithms from Anagram of Switzerland. It looks like they are doing ASRC (Asynchronous Sample Rate Conversion) of all sources to 192/24 and then feeding it to the Wolfson DAC chip."
That's Anagram's product, sample-rate conversion. ASRC as far as I'm concerned is a red flag for any real-time digital playback..... It's touted as a jitter reducer, but in reality it transforms the jitter to noise.
If the filter functions are controlled at the Wolfson DAC, the Anagram SRC ahead of it would override the functionality. Once the ASRC is performed, the filtering from the SRC module is already done, and cannot be reversed or re-done at the DAC. (Unless the minimum phase option bypasses the SRC.)
"I would assume that the Wolfson further oversamples it another 8x, but it is not clear if this is the case or not."
Most oversampling DAC chips only do so for Fs at 48 kHz and below. At least the data sheets I've looked at suggest that.
"Most oversampling DAC chips only do so for Fs at 48 kHz and below. At least the data sheets I've looked at suggest that."
Looking at the Wolfson 8740 data sheet, this DAC apparently oversamples for inputs all the way up to 192 kHz. The only thing this would do is enable use of a smaller analog post filter..... I personally think digital filtering of 24/96 or 24/192 kHz signals is unnecessary, and probably detrimental. Likely more RFI generated from the extra processing.
I hope the designers of the DACMagic didn't design the filter options controlled from the Wolfson DAC..... With the ASRC feeding the DAC a 24/192 signal, the filter options wouldn't be acting on the 16/44 signal off the CD, but the 24/192 signal fed from the ASRC. (The differences in the filter options would be virtually inaudible running this way.)
Or more correctly, most of the digital filtering is bypassed. There is still a gentle wideband filter applied. And all of the audio digital filters (or DAC in this case since the DF is in the same package as the D/A convertor) I'm aware of oversample the input regardless of the sample rate, that is what the maximum sample rate specification is for. One could control externally whether the input bypasses the digital filter at higher rates, and that is done by some, but it isn't commonplace when using separate digital filter and D/A convertor parts because it requires extra logic to get the signals in the right format if you bypass the DF. On the Wolfson, it is via an extra data input since the signals need to be in the correct format with separate left and right data.
Edits: 02/11/09
I guess the filter function is controlled at the ASRC upsampler, which would be the place to do it. (But that does not change my opinion of ASRC.) The DACMagic literature is a little confusing, because the filter control is brought up in the "Wolfson DAC" section.
The upsampler must be a proprietary design, since I'm not aware of an off-the-shelf SRC chip with these filter options.
Charles H said the code was from Anagram, but not sure where he got that info. Might show up in a google.I think the Wolfson WM8741 parts are also being used in Linn's flagship players, but again, don't know how many of the filter options are available to the user. I'd assume all of them in such expensive players. I think some of the models, like the computer interface model, use their own digital filter, though. Maybe the Wolfson parts aren't really that good, but strange they would use them just for the DAC section.
Edits: 02/11/09
Hi-Fi News gave this product a rave review. That's where I saw the information about the collaboration with Anagram.
If you do a search of the archives here, you will find that many owners have been less enthusiastic about the sound than the reviewer was. In all fairness, the product is very low priced -- I think the UK price was 250 pounds. It's hard to make something good for that much money, even when it is made in China.
One thing that was surprising was that how very small changes to the filter parameters could have relatively big sonic impacts.
You can download a demo version of the filter program we used and play around with things. (The full version is $2,000.) But until you actually *listen* to the filters, you don't know what sounds best. We were pretty methodical about it and it still took us months to arrive at our final conclusions.
> Years ago anybody trying to implement their own digital filter
> either had to design a custom chip (very expensive) or write
> software for DSP (digital signal processor), less expensive
> but inefficient.
There was also (back in '93, was it?) Pioneer's Legato Link
chip (used in the Elite PD-65 player), which was similar enough
to what Wadia was doing in DSP software that some kind
of legal tussle ensued.
Pioneer has done some cool stuff, over the years, for a
mass-market manufacturer.
I suspect that the Burr-Brown "slow roll-off" filters were the result of requests from Pioneer. They are about the only company big enough to have their requests heard, and they are also about the only big company with an interest in anything besides brickwall filters.
The problem with FPGAs is one would need training in how to program and implement the things..... I personally have never had experience with them......
"Then there is the issue of do you use the same filter for all sample rates?"
The filter is only needed for lower sampling rates..... If it's over 60 kHz, I think an analog RC filter is sufficient. The problem is below that rate, there are modulation and ringing issues in regard to the uppermost part of the 20Hz-20kHz audio band...... I don't think 96 kHz playback needs digital filtering. Just use an RC network that allows full amplitude at and below 20 kHz and full attenuation at and above 48 kHz.....
"The big tradeoffs are for 44.1 since it is so close to the audio band."
That's the sole reason why digital filtering is utilized for 44.1 kHz playback.....
"Does it make sense to use the same filter at 96 or 192? maybe not. I don't think there has been a lot of exploration of this aspect."
I don't think 96 and 192 capable DACs apply filtering for settings above 48 kHz. (88.2 kHz, 96 kHz, 176.4 kHz, and 192 kHz.)
> The problem is that we don't really know exactly what to optimize for.
> There is no equation that will design a filter that "sounds best". . .
And then of course there was the Pacific Microsonics HDCD approach --
have the A/D converter "decide", based on the incoming signal, which
of a number of playback filters would sound best, and then put control
bits in the data stream directing the playback chip to switch filters
dynamically.
But it's never been clear to me that the D/A filter also changed. There is absolutely no hint of any such thing happening in spec sheets for either the PMD-100 or the PMD-200.
And when you read the white papers, they *very* carefully dance around this issue.
I would love to hear from Keith Johnson and put this issue to rest for once and all.
I wrote for the fun my own HDCD software decoder.
I could tell you, that on the few HDCD encoded disk I tried (essentialy Neil Young ones), there is what appears to be switching filter commands but
only between 2 filters.
The "alternate" filter seems to be used on high frequency contents.
I could give you more precise example if you want....
It doesn't means that PMD-x00 chip do something with these commands.
and it's true that PMD spec sheets don't tell too much.
PCM1732 spec don't tell really more , but a least it shows a link between HDCD decoding and digital filter....
Killer project!
I saw some other forums that mentioned a project like this, was this your decoder or are there two smart people out there with that much time on their hands?
I'm not sure how you would know that the commands being sent are to change the filter, as opposed to the other parameters that can be changed.
This might be better to take off-line, so please e-mail me at "chansen at ayre dot you-know-what" and be sure to put the word "Ayre" somewhere in the subject line of your first e-mail so that it gets past my spam filter. After that I can add you to the "approved list".
> I saw some other forums that mentioned a project like this, was this your > decoder or are there two smart people out there with that much time on > their hands?
No the same one
> I'm not sure how you would know that the commands being sent are to change > the filter, as opposed to the other parameters that can be changed.
You have a mail
Wow....thats great Charles
Looks like that digital filter is what was needed after all....I guess there 'are' more ways than one to skin an audio cat.
Karma Means Never Having To Say You're Sorry
There were so many great things said for so long about the sound of "non-oversampling" (filterless) DACs that I was very curious. I think that part of the results may be dependent on the architecture of the DAC itself, and also the system into which it is placed.
By having no digital filter, the analog output is a "stair-step" signal. This can be smoothed with analog filtering. Many NOS DAC designs have transformers in the analog stages, which tend to roll off the ultrasonic frequencies quite rapidly.
Even then there will be "aliasing" and other high-frequency artifacts in the signal that will disturb some preamps and amps more than others. So different people may reach different conclusions, especially considering the wide variety of systems in use.
I want to ask how can your filter distinguish between music at its set frequency and artifacts? Invariably they will be intermingled together, I saw your white paper and it reads/looks very impressive. I have heard your products at a store in Michigan [Audio Dimensions - owner was Harry] and I must say I did like them a lot. So I do tend to believe this implementation will in fact deliver good sound [I trust your designs].
Karma Means Never Having To Say You're Sorry
I'm not sure that I understand your question. No filter can distinguish between music and "artifacts". I'm guessing that I wrote something in the white paper in way that was confusing.
I'm looking forward to auditioning the technology. I think this will push the state of the art for CD (or CD quality) playback.
I think it made a big difference. But more importantly it will be good to get reports from users and reviewers.
With my tube DAC, things improve even more so when I disconnect the Oppo DVD player or Marantz CD63SE and use the Denon DCD-1650Ar as a transport - more so when I use a coaxial rather than an optical cable.
...Just in case you wanted to experiment with transports.
The Denon DCD-1650AR is one heavy duty and well built CD player. And it sounds pretty darn good too! I had one of those many years ago and also had the Sony XA7ES. Nice gear.
Those were the days.
Its in my bedroom system, getting daily use.
Jack
.
I've had the Alpha, too, and your reaction was exactly how I reacted in the beginning.
I would say give it time and see if you can live with its shortcomings. At first the tube color, richness, bloom, 3-D atmosphere can be a fresh breath of air, especially if you're coming from poorly-designed (read: cheap opamp output stage) SS DAC's or CDP's.
Ultimately, I needed much more detail resolution, speed, attack, and mostly bass definition, which I found in a nice DAC with discrete SS output stage... Then again, maybe your tastes will be fine with the Alpha.
Post a Followup:
FAQ |
Post a Message! |
Forgot Password? |
|
||||||||||||||
|
This post is made possible by the generous support of people like you and our sponsors: