Skip to main content
Topic: upsampling/downsampling artifacts (Read 7336 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

upsampling/downsampling artifacts

I'm wondering if repeatedly upsampling and downsampling digital audio could lead to (cumulative) interpolation/decimation filter artifacts becoming audible, or cumulative filter response audibly affecting frequency response below 20kHz?

Anyone can help me out here (or point me in the right direction)?

Thanks

drewfx

upsampling/downsampling artifacts

Reply #1
Well, if the attenuation at 15khz for a particular reconstruction filter is 0.01db for a downsampling filter, then it should "only" take 100 downsamples for that to hit 1db.....

upsampling/downsampling artifacts

Reply #2
Quantization noise will be added with each upsample/downsample, but this noise only increases as the square root of the number of conversions.

upsampling/downsampling artifacts

Reply #3
It probably depends on the quality of the filters
A few months ago I've done some testing by upsampling&downsampling a 32-bit fp signal 10 times from 44.1-88.2kHz rate using iZotope's SRC.
Then I subtracted the resulting signal from the original. The only difference I could see (not hear) was some artifacts above 21kHz. Below that there was nothing to see, indicating to me that the process is kinda lossless for source material up to 21kHz.
Nevertheless there are many claims about (audible) effects of SRC but I doubt that they can be confirmed by double blind tests, assuming the SRC is of good quality of course.

upsampling/downsampling artifacts

Reply #4
I just wanted to point out that while repeated upsample&downsample is an excellent way to amplify the effects of a single upsample or downsample, the same can not be said for lossy compression.

From time to time someone comes here with the resuts of repeated lossy compression using multiple codecs, and tries to draw a conclusion from that.

The effects of upsampling and downsampling are additive in a fairly simple way, but the interactions of successive stages of lossy compression are much more complex and do not accurately indicate the effect of a single compression.

upsampling/downsampling artifacts

Reply #5
Two years ago we created a 20th generation A/D - D/A dub where each A/D and D/A conversion used and AD1896 sample rate converter.  There were a total of 40 ASRC operations.  We found some material where we could hear the differences in an ABX test (original vs. 20th generation), but we determined that the differences were due to a slight high-frequency roll-off in the A/D and D/A analog filters and were not due to any artifacts caused by the ASRC process.  We confirmed this by correcting the frequency response in a DAW and then repeated the ABX tests.  After the minor frequency response correction, we could not hear the differences in the ABX tests. 

We conducted the tests in our listening room and at Avatar Studios in NYC.  The A/D was a Benchmark ADC1, and the D/A was a Benchmark DAC1.  The subjects were audio professionals.  The audio material was recorded at 96 kHz, 24-bits.  The ADC1 performed a 96kHz to 96 kHz ASRC operation while the DAC1 performed a 96 kHz to 110 kHz ASRC operation.  We used a hardware ABX test box that uses relay switching, 20 trials per session.

If there is interest, I could make the files available to this forum.

The AD1896 has distortion artifacts that are less than -135 dBFS (well below audibility thresholds).  Some newer ASRC devices are even better.  The quality of an ASRC process is only limited by the amount of DSP processing that we are willing to expend.

Two words of caution: 
1) The ASRC processing that is built into many computer operating systems is very poor.  Artifacts can be as high as -70 dBFS.  Windows Vista is the first operating system that has decent sample rate conversion. 
2) ASRC processes can add jitter-induced distortion if the ASRC rate-estimator does not have sufficient jitter attenuation.  The AD1896 is one of only a few processors that have good jitter attenuation.
John Siau
Vice President
Benchmark Media Systems, Inc.

upsampling/downsampling artifacts

Reply #6
John_Siau, that's interesting and reassuring.

However I see the SR used was 96KHz.  I imagine that a sample rate of 48KHz, and to an even greater degree, 44.1KHz, would have been more susceptible to cumulative high-frequency roll-off.  A common SR for Blu-ray discs is 48KHz.

upsampling/downsampling artifacts

Reply #7
John_Siau, that's interesting and reassuring.

However I see the SR used was 96KHz.  I imagine that a sample rate of 48KHz, and to an even greater degree, 44.1KHz, would have been more susceptible to cumulative high-frequency roll-off.  A common SR for Blu-ray discs is 48KHz.

Yes, I doubt that we would have had the same results at 48 kHz or 44.1 kHz.  At 96 kHz the filters are much less apt to cause problems.

Also, there are major variations in the quality of sample rate conversion.  Distortion in some devices is as high as -70 dB FS while others are better than -140 dB FS.

Generic statements about the audibility (or inaudibility) of sample rate conversion should be avoided.
John Siau
Vice President
Benchmark Media Systems, Inc.

upsampling/downsampling artifacts

Reply #8
Generic statements about the audibility (or inaudibility) of sample rate conversion should be avoided.
Agreed. But IMHO anecdotal experiences should be labeled as such
It's probably also a good idea to distinguish between real-time SRC (like in hardware) and non-realtime applications (file processing).

upsampling/downsampling artifacts

Reply #9
I can’t say about audibility but there are certainly striking artifacts from some software, so the answer to the question may well depend on which software one uses. Compare Adobe Audition using the pre/post filter with Goldwave, Albeton Live,  SoundForge, no anti-aliasing filter, or Wavosaur 1.0.1.0 on the sweep frequency conversion.

http://src.infinitewave.ca/

upsampling/downsampling artifacts

Reply #10
Thanks for all of the input.

I'm wasn't really thinking about the quality of one SRC vs. another (that's a separate subject), but rather whether inaudible artifacts could become audible after several iterations. I was thinking about serial use of (increasingly common) upsampling/downsampling (VST) processes in a DAW.

It occurred to me that you might conceivably pay a penalty due to cumulative filtering effects, and thus there may actually be an audible benefit to running the host at a higher sampling rate at some point.

drewfx

 
SimplePortal 1.0.0 RC1 © 2008-2019