If for a continous analog output signal, what is the filter interpolating between? And I thought clocking was in the digital realm to minimize signal bit timing synchronizing jitter.
I am confused as well, interpolation means between discrete points, can only be done in the digital domain?!
Otherwise, @judmarc: I have had good experience with NOS / analog reconstruction "brickwall" filter DAC and I typically dislike oversampling it seems (although my current DAC may actually be oversampling 8x for all I know yet it does not suck to these ears). So maybe there's more to it than the R2R / NOS buzzwords you mention indeed.
My thoughts on the topic, posted earlier in the stax thread (
post #3960 )but more relevant here actually :
Why do we limit the frequencies? Well, human ears can't hear sound over a certain frequency, so we just cut it off.
To answer simply: it's not quite that simple I am afraid.
I stand to be corrected with what's below as I am writing from memory and memory can be treacherous sometimes but...
- Indeed with a continuous fourier transform (and its inverse), we can go to and from time or frequency domain representations of a signal. This assumes you do not truncate any piece of information in either domains
- Issue is when we go from continuous to discrete (or analog to digital domains if you will), as you run the risk of truncating some of the frequency content when doing the finite fourier transform
- Without filtering of this high frequency content (basically anything above 1/2 sample rate), it will "fold back" onto the visible portion of the fourier spectrum.
- Without even considering fourier transforms and sticking with the conversion of time domains signals from contiguous to discrete domains, the above explains why "anti-aliasing" filters are necessary prior to an A/D stage.
- Similarly, I believe "reconstruction filters" are necessary after D/A stage to smooth out (filter out) high frequency noise that results from the quantization process. e.g. a stair case has effective very broad frequency content well above the sinusoidal it is trying to replicate.
Now where that relates to practice:
- Astrostar is trying to share his bad experiences with oversampling and other reconstruction filters used in D/A stages.
- For instance, no reconstruction filter is gentle if effective, hence the nasty name "brickwall".
- The steeper the slope of the filter (required if you want to maintain bandwidth right up to half the sample rate), the nastier it's transients (e.g. you get lots of pre-ringing or lots of post-ringing or some mix along with phase distortion)
- The solution has long been to either a) get rid of the reconstruction filter (that assumes you trust the high frequency hash you may get as a result isn't going to bite you back), b) oversample the data so that the filter can have gentler slope, c) forget about trying to reproduce 20kHz tone without attenuation and leave with a slight HF roll off from the filter at the benefit of reduced ringing.
- You'd think oversample is the most elegant way to deal with the issue but it turns out an oversampling filter is just another filter with its own issues in terms of transient response and such.
My personal experience:
- I've heard nasty artifacts with a NOS/filterless DAC I tried a few years back so typically running away from filterless designs.
- I've heard no better than NOS / R2R DAC (with a brickwall filter).
- I've never ever preferred a resampled / up sampled version of a 44.1k, be it software or hardware based upsampling
- My take is that oversampling is thus more detrimental than a well designed NOS reconstruction filter and that is the direction I am headed to.
cheers,
arnaud