Nika Aldrich wrote on Tue, 26 October 2004 08:29 |
Joe Bryan wrote on Mon, 25 October 2004 06:38 | In many cases, this makes it impossible to reproduce the exact transfer function without upsampling. For example, it's impossible to recreate the exact amplitude and phase characteristics of most high-order (>2) analog filters in digital without upsampling. A perfect example is the Pultec EQ. Any DSP process that claims to be a Pultec but doesn't upsample isn't matching the response correctly.
|
There are a lot of EQs out there that successfully do proper DSP without upsampling. The Sony Oxford is a classic example.
|
Yes, there are plenty of EQs that can produce good filter responses without upsampling, but that does not mean they can produce every desireable filter response. As I said, there are analog filters that cannot be reproduced digitally without upsampling.
Nika Aldrich |
Quote: | I was referring to the analog filters, not the digital filters.
The digital filters only do part of the work. For a typical sigma-delta ADC, the analog anti-aliasing filter is trivial because it only needs to attenuate signals above the MHz range.
However, the analog anti-imaging filter for a DAC is not trivial. In today's sigma-delta converters, there is a lot of out-of-band energy that cannot be removed by the digital filter because it's ahead of the signal-delta DAC's modulator. The modulator's noise shaping shifts the quantization noise out of the main audio band, but this energy can only be removed by the analog filter.
|
Sure, but the roll-off of that analog filter is so far out of band that increasing it further is unnecessary with respect to phase shift in the audible range.
|
It's not "so far out" as you say. The one plot you won't see on any of the DAC spec sheets is the out of band noise caused by the modulator's noise shaping. This rises quite rapidly above the digital filter's stop band, and requires some non trivial analog filtering to remove it. The analog filter must be set as low as possible to remove this noise, and this requires juggling greater in-band phase error vs. greater out-of-band noise levels.
Nika Aldrich |
Quote: | It is this filter that benefits the most from higher sampling rates. The higher transition band provided by higher sampling rates allows for much less phase shift in the primary audio band while attenuating the high-freq energy that wreaks havock in the analog output amps. This has a major impact on sound quality, especially transparency and transient response.
|
On a typical DSM based DAC the oversampling brings the excess noise so far above the audible range that a very gentle filter can be used, just like on the A/D side. I see no grounds for claiming that making the filter even more gentle is better on audio band material - especially if the current system provides no audible phase distortion because the filter is already gentle enough.
|
See above. Unless my ears and test equipment have been deceiving me all these years, it does make a difference.
Nika Aldrich |
Quote: | I can't think of any examples where this is true unless you're referring to up/down SRC when the processing is oversampled but not the analog conversion, could you provide some?
|
There are plugins that, when run at higher sampling rates, have to down sample in order to do the processing and then upsample at the end of the process.
|
Ah, yes. Forgot about those. We don't usually have to deal with that problem because we still haven't run out of cycles on our DSP.
Sometimes downsampling is done when the process doesn't benefit from the higher audio bandwidth, and the excess cycles are unneccessary. Digital reverb is one common example, and because of the time-smearing nature of reverb, the added delays from the SRC are trivial. This type of reverb would still benefit from lower-latency, high sample rate processing in a monitoring environment as long as the direct path wasn't downsampled.
Another example is decimated linear phase filters, which have high intrinsic delays (like any linear phase process). These would never be used in a monitoring situation to begin with.
The only other case where downsampling is necessary is when a process can't run at full speed without overloading the system. No algo designer wants to do this, they're forced to do it to get the process to run, even if it compromises quality.
-Joe