Thanks Don,
Well it was to get your reaction to Seokris comment;
"imagine the digital tricks you need to do to get music out of a 1-5 bit DAC"
Well yes I can imagine the complexities and understand them to a limited level, but I'm not a DAC designer. His implication is that the DS technique leads to unsurmountable problems that make all DS designs significantly flawed for audio. Do you feel this is the case?
Oh BTW there appear to be amipro converters, found this in a quick search
http://file-convert.com/a_ami.htm
This will be hand-waving and bear in mind I am not a "real" delta-sigma expert. I have designed a few, and been taught by a few experts (including Dr. Gabor Temes at UCLA, great guy, nice and brilliant), but most of my designs were more conventional designs. My career was mostly high-speed and oversampling was difficult when you were already pushing the process technology limits. Except for "fun" I have not designed an audio DS data converter.
Noise modulation, essentially varying the noise floor with the signal, is an oft-cited concern with DS designs. However, a small amount of dither usually solves that issue, especially for high-order and multibit loops, since they are easier to to decorrelate. Limit cycles happen when essentially the same samples pass through the filter and so spurs "bunch up" at discrete frequencies (tones). Idle tones are discrete signals that stick out well above the noise floor when the input is essentially static (DC, mentioned in my previous post). Harmonic distortion can get tricky; the loop and filters have been shown to introduce harmonic distortion that is not what I would call "classical" nonlinearities, and I do not claim to fully understand it. It's been a while. Low-order, single-bit DS circuits are more susceptible to these errors than higher-order designs that tend to "randomize" the errors more. IMO this is due to both the greater complexity (more things going on means less chance any one thing is going to "add up") and the fact that with more circuits are more noise sources to randomize (decorrelate) spurs.
One thing that sometimes gets overlooked is that even a 1-bit delta-sigma requires full precision at some point in the circuit. For an ADC, that is the difference (delta) circuit at the input; errors there will not be compensated through the rest of the loop. If you want a 16-bit answer, you must accurately difference the feedback to 16 bits. For multibit designs, the first (at least) feedback DAC must be accurate to the full number of bits. That is, if you have a DS ADC with a 5-bit DAC in the feedback loop, those 32 steps must be accurate on 16-bit boundaries (one part in 65,536) to achieve 16-bit precision at the output. Most often there is a large digital engine that calibrates and compensates the loop components to achieve high precision, either by directly modifying the individual transfer functions, or by essentially post-processing (on the fly) the output and adjusting the samples based upon a calibration routine. That is mostly outside my scope; I had help with the digital engine afterwards, and simply used Matlab to create the compensatory weighting when I tested my modulators (the compensating engine was a DSP, FPGA, or custom high-speed logic at various times and for various applications).
I do not claim to have golden ears, more like ears of clay, so hesitate to say delta-sigma designs are fundamentally flawed for audio. The sheer number of them in use, the huge number of papers published showing their performance under all sorts of conditions, and the continuing advances in their architecture and design makes me think not. I have much less experience with delta-sigma DACs, however, and do not claim to be any sort of expert in their design. I am well aware of the drawbacks of conventional converters and delineated some of them earlier. Their issues are by and large understood and more straight-forward (I would not say "easier") to analyze though compensation can still be tricky. When you are looking >100 dB below full-scale
everything matters and it is easy to let some little thing slip through.
For example, resistor and/or transistor self-heating is a big issue at high resolution. Think of a ladder DAC, R-2R or unary, and how the signal interacts along the ladder. At the top and bottom, the switches rarely change state since those codes are exercised infrequently, so the resistors and switching transistors tend to be in one position always. Since current flows (however small) mainly through one side, it runs hotter than the other side. Now, even if the switch and resistors were perfectly matched (or more likely trimmed) to begin with, this means they will have slightly different gain and offset when they switch to the other state (every now and then). In the center of the ladder, around the zero-crossing of the signal, both sides are used all the time so both sides of the switch and resistors stay essentially the same temperature. In the end you have a thermal bow related to the signal -- a square wave will distort differently than a sine wave. Ain't that nice.
Anyway, chances are I've lost most of the audience here already, so I'll quit here. I do not necessarily disagree with
@soekris, but conventional converters have their quirks as well (even sign-magnitude, which does have some nice features to get around some of the issues).
This old AES paper is online and worth a look for those interested in a deeper, though still fairly high-level, look at delta-sigma designs:
http://www.eecs.qmul.ac.uk/~josh/do...maDeltaModulation-SolvedandUnsolvedIssues.pdf
HTH - Don
p.s. I suspect part of the reason some prefer conventional DACs is because they typically have low-order distortion characteristics, and without noise shaping have a flatter noise floor, sort of like tube circuits that have measurably higher distortion but often sound better to audiophiles.
