• WANTED: Happy members who like to discuss audio and other topics related to our interest. Desire to learn and share knowledge of science required. There are many reviews of audio hardware and expert members to help answer your questions. Click here to have your audio equipment measured for free!

DSP Measurements and Rising Noise Floor

If you scan a library of lossless CD tracks with an upsampling reconstruction filter, you will find that most overs are less than +2 dBFS. I haven't seen any over +3 dBFS, and I'm not entirely sure that anything above 3.01 dBFS is possible when the input signal has been properly band limited. In any event, 3 dB of headroom is sufficient.

We used 3.5 dB in the DAC2 converters and then backed this down to 3.0 dB in the DAC3 converters because we concluded that 3.5 dB was overkill. Every dB reserved for intersample overs reduces the SNR of the DAC by 1 dB, so you don't want to reserve more than is necessary. I have not found a CD recording with an over that exceeds +3.0 dBFS, but one might exist.
See this thread.
 
I think your link is incorrect, it takes me to a product page for the HPA4 firmware upgrade kit. I imagine you meant this one -> https://benchmarkmedia.com/blogs/application_notes/intersample-overs-in-cd-recordings.

Good points on intersample overs, especially as it relates to "transparency" claims, as you note the ASRC in the miniDSPs can definitely have issues with this. However, to me that is a pretty well known issue and was not the point of these tests. As you note I specifically chose lower levels to avoid this. What I wanted to learn more about was the potential for elevated noise / distortion when applying filters.

Michael
Thanks Michael, I corrected the link in the first post.

Your work is tremendously useful, but I'm curious to know how many other DSP boxes have the same problem as the miniDSP. Some additional test above -1 dBFS would expose these problems, if they exist.

The problem with the miniDSP is that you cannot turn off the ASRC nor can you reduce the input gain in front of the ASRC. If you reduce the input signal before sending it to the miniDSP it works well.

We frequently use the miniDSP SHD devices in our listening tests because they are a versatile DSP platform, and I think we have about six SHD units here. Unfortunately, many end users will not be aware of the need to reduce the signal level in the digital domain before the miniDSP. They may also lack the necessary equipment (or software) to do this.
 
Thanks Michael, I corrected the link in the first post.

Your work is tremendously useful, but I'm curious to know how many other DSP boxes have the same problem as the miniDSP. Some additional test above -1 dBFS would expose these problems, if they exist.

The problem with the miniDSP is that you cannot turn off the ASRC nor can you reduce the input gain in front of the ASRC. If you reduce the input signal before sending it to the miniDSP it works well.

We frequently use the miniDSP SHD devices in our listening tests because they are a versatile DSP platform, and I think we have about six SHD units here. Unfortunately, many end users will not be aware of the need to reduce the signal level in the digital domain before the miniDSP. They may also lack the necessary equipment (or software) to do this.

Do you have a +3 dBFS 11.025 kHz test tone you can share?

I can probably generate one in Audacity but I am not that proficient in it and want to make sure if I am testing something that I am not starting from a bad test tone. I did try a 44.1 kHz sample rate, 11.025 kHz test tone at 0 dBFS from REW in to the miniSHARC with no filters applied and it was clean, although I think that is somewhat expected.

Michael
 
Do you have a +3 dBFS 11.025 kHz test tone you can share?

I can probably generate one in Audacity but I am not that proficient in it and want to make sure if I am testing something that I am not starting from a bad test tone. I did try a 44.1 kHz sample rate, 11.025 kHz test tone at 0 dBFS from REW in to the miniSHARC with no filters applied and it was clean, although I think that is somewhat expected.

Michael
It is not difficult to construct. Here is an example of a +10 dB over.

intersamples_over.png
 
Do you have a +3 dBFS 11.025 kHz test tone you can share?

I can probably generate one in Audacity but I am not that proficient in it and want to make sure if I am testing something that I am not starting from a bad test tone. I did try a 44.1 kHz sample rate, 11.025 kHz test tone at 0 dBFS from REW in to the miniSHARC with no filters applied and it was clean, although I think that is somewhat expected.

Michael
Here you go (exported from Multitone,tell me if you want different sample rate,duration,etc)

 
Do you have a +3 dBFS 11.025 kHz test tone you can share?

I can probably generate one in Audacity but I am not that proficient in it and want to make sure if I am testing something that I am not starting from a bad test tone. I did try a 44.1 kHz sample rate, 11.025 kHz test tone at 0 dBFS from REW in to the miniSHARC with no filters applied and it was clean, although I think that is somewhat expected.

Michael
And here is couple more to test,they can really cause mayhem in some DACs.


(thanks to @pkane for the tests)
 
If you scan a library of lossless CD tracks with an upsampling reconstruction filter, you will find that most overs are less than +2 dBFS. I haven't seen any over +3 dBFS, and I'm not entirely sure that anything above 3.01 dBFS is possible when the input signal has been properly band limited. In any event, 3 dB of headroom is sufficient.

We used 3.5 dB in the DAC2 converters and then backed this down to 3.0 dB in the DAC3 converters because we concluded that 3.5 dB was overkill. Every dB reserved for intersample overs reduces the SNR of the DAC by 1 dB, so you don't want to reserve more than is necessary. I have not found a CD recording with an over that exceeds +3.0 dBFS, but one might exist.

If you want a pathological example, pick up Merzbow's Venereology. Most tracks have more than 7dBFS of inter-sample x-overs with a typical reconstruction filter, and one hits 11.30dBFS.
 
Here you go (exported from Multitone,tell me if you want different sample rate,duration,etc)


Thank you!

Here is TOSLINK out at 44.1 kHz into the UR23 running at 44.1 kHz. Looks perfectly clean at +3 dBFS.

TOSLINK to UR23 intersample over.png


And now that same 44.1 TOSLINK output in to the miniSHARC with -10 dB input gain applied in attempt to eliminate clipping and capture by the UR23 at 96 kHz. Obviously looks really bad.

miniSHARC intersample over.png


To me even more reason to use CamillaDSP as you can capture with something like the UR23 and then drop the level by a few dB in CamillaDSP to safely handle intersample overs.

Michael
 
Do you have a +3 dBFS 11.025 kHz test tone you can share?

I can probably generate one in Audacity but I am not that proficient in it and want to make sure if I am testing something that I am not starting from a bad test tone. I did try a 44.1 kHz sample rate, 11.025 kHz test tone at 0 dBFS from REW in to the miniSHARC with no filters applied and it was clean, although I think that is somewhat expected.

Michael
The attached file is an 11.025 kHz tone at 45 degrees relative to the samples, normalized to 0 dBFS. This creates an intersample peak that is +3.01 dBFS. Use extreme caution when playing this file. It can easily take out a set of tweeters!
 

Attachments

  • Caution-HighLevel11025Hz-IntersampleOverTestFile.zip
    374.8 KB · Views: 80
I have not found a CD recording with an over that exceeds +3.0 dBFS, but one might exist.
I in fact own several.

The worst offender here is actually on a fairly common album, Santogold s/t from 2008 (you may remember the singles like L.E.S. Artistes or Say Aha) - it's track #12, the You'll Find A Way remix. RG scanning with SoX oversampling to 176.4 kHz reveals a peak amplitude of 1.648287. I imagine the used CD bins ought to yield a copy fairly easily.
Next up, the Sneaky Sound System (European) comp from 2009. Worst offenders are I Want Everything (peak 1.608418) and It's Not My Problem (peak 1.586710).
The last one over +3dB is found in another common album, Grimes' Art Angels from 2015. Track #7, Artangels peaks at 1.483201.

That's out of 8300 tracks and change.
 
Imo the trouble with intersample overs is that they are a sound engineer mistake (afaik). Trying to fix that in the dac creates more room for more and higher intersample overs.

A rough estimate from the few samples given here is that it occurs somewhere between 1 in 1000 and 1 in 2000 samples. The effect would be audible in higher frequencies.

A remedy would be to digitally reduce the volume in our digital players by about -3db to cure most occurences. Or to do so in the dac. The tradeoff will be lower sinad at the output.

All the above makes it a valid finding, but imo with a minor risc and small audible impact. Ymmv, everyone should decide for themselves.

Kudos to Benchmark for being so thorough in their testing and engineering.
 
Last edited:
Thank you for a very interesting and valuable work!

I have read this and the great articles from benchmarkmedia to better understand the problem with intersample overs.

@John_Siau at your webpage (in Q&A section) you are addressing the problem with intersampling overs by not recommending using a software approach:

Q: Many folks use software (HQPlayer, Roon, etc.) based sample rate and/or format (PCM<->DSD) conversion upstream of the DAC. Do you see any benefit to this, and if so, is there a preferred sample rate and/or format for your DACs?

We do not recommend upsampling fixed-point digital audio as this will clip intersample peaks. We do this upsampling inside of the DAC2 and DAC3, but we do it with adequate headroom so that clipping cannot occur.

If I have understood this correct, does not this test show that interpolating upstream like CamillaDSP does it, is indeed a very good way of doing it instead of the dac? To quote @mdsimon2:

To me even more reason to use CamillaDSP as you can capture with something like the UR23 and then drop the level by a few dB in CamillaDSP to safely handle intersample overs.

This problem is new to me, so please bear with me if I have misunderstood something.
 
Last edited:
I in fact own several.

The worst offender here is actually on a fairly common album, Santogold s/t from 2008 (you may remember the singles like L.E.S. Artistes or Say Aha) - it's track #12, the You'll Find A Way remix. RG scanning with SoX oversampling to 176.4 kHz reveals a peak amplitude of 1.648287. I imagine the used CD bins ought to yield a copy fairly easily.
Next up, the Sneaky Sound System (European) comp from 2009. Worst offenders are I Want Everything (peak 1.608418) and It's Not My Problem (peak 1.586710).
The last one over +3dB is found in another common album, Grimes' Art Angels from 2015. Track #7, Artangels peaks at 1.483201.

That's out of 8300 tracks and change.
Thanks for these, I will take a look at them.

Of the 8300 tracks, how many have peaks that exceed +3.0 dBFS? Just these three, or do you have more?

I am trying to determine if our allowance of 3 dB for intersample peaks (at our calibrated volume position) is the best tradeoff between DSP headroom and SNR. We have more headroom at lower volume settings.
 
Thank you for a very interesting and valuable work!

I have read this and the great articles from benchmarkmedia to better understand the problem with intersample overs.

@John_Siau at your webpage (in Q&A section) you are addressing the problem with intersampling overs by not recommending using a software approach:

Q: Many folks use software (HQPlayer, Roon, etc.) based sample rate and/or format (PCM<->DSD) conversion upstream of the DAC. Do you see any benefit to this, and if so, is there a preferred sample rate and/or format for your DACs?

We do not recommend upsampling fixed-point digital audio as this will clip intersample peaks. We do this upsampling inside of the DAC2 and DAC3, but we do it with adequate headroom so that clipping cannot occur.

If I have understood this correct, does not this test show that interpolating upstream like CamillaDSP does it, is indeed a very good way of doing it instead of the dac? To quote @mdsimon2:

To me even more reason to use CamillaDSP as you can capture with something like the UR23 and then drop the level by a few dB in CamillaDSP to safely handle intersample overs.

This problem is new to me, so please bear with me if I have misunderstood something.
We use Roon to apply a 3 dB reduction prior to the miniDSP SHD processor whenever we use that processor in our listening rooms. We do not apply this 3 dB reduction when feeding the Benchmark DAC3 directly from Roon. We have done the same 3 dB reduction using JRiver and several other high-end players. We have found that the DSP headroom functions perform as advertised.

DSD does avoid the intersample clipping problem if the DSD is converted natively by the D/A converter. Conversion from PMC to DSD is not the solution to the PCM intersample over problem. This PCM to DSD conversion process may introduce clipping of the intersample peaks if the conversion processor lacks sufficient headroom. The correct solution is to apply a 3 dB reduction before the first DSP process (including the interpolation in an oversampled sigma-delta D/A converter chip). The output of the 3 dB reduction should be 24 bits and not 16 bits.

I would never recommend converting PCM to DSD. This is a lossy process that always adds noise and distortion to the audio. This loss in quality is easy to predict mathematically and it is easy to measure. But, the reduction in performance is small enough that it may not be noticeable. The performance of DSD64 is very good, but it is not as good as 44.1/24.

DSD is a terrible format for studio production, and it is problematic for the end-user who would like to add a DSP process such as a volume control, a crossover, or a room correction system.

I would recommend converting DSD to PCM when you wish to add a DSP process such as a crossover, EQ, or room correction. The DSD to PCM conversion process is mathematically transparent except for the reduction in bandwidth. However, the reduction in bandwidth may be advantageous because it removes the high-amplitude ultrasonic noise from the DSD source. With DSD64, any musical content above 22 kHz is obscured by the DSD noise shaping, so the DSD bandwidth above 22 kHz is not useful because of the high noise levels. Some power amplifiers and tweeters can fold this ultrasonic noise into the audible band, so conversion to PCM may actually improve the performance at the output of the loudspeakers.

Convert DSD64 to 44.1/24. Convert DSD128 and higher to 88.2/24. Reduce the PCM level by 3 dB, then apply your DSP processing.
 
Imo the trouble with intersample overs is that they are a sound engineer mistake (afaik). Trying to fix that in the dac creates more room for more and higher intersample overs.

A rough estimate from the few samples given here is that it occurs somewhere between 1 in 1000 and 1 in 2000 samples. The effect would be audible in higher frequencies.

A remedy would be to digitally reduce the volume in our digital players by about -3db to cure most occurences. Or to do so in the dac. The tradeoff will be lower sinad at the output.

All the above makes it a valid finding, but imo with a minor risc and small audible impact. Ymmv, everyone should decide for themselves.

Kudos to Benchmark for being so thorough in their testing and engineering.

These are all really good points.

After experimenting it seems intersample overs are really only an issue where you have SRC without prior attenuation. Every miniDSP product tested in this thread does not cleanly pass the +3 dBFS 11.025 kHz tone, regardless of miniDSP input gain or volume control setting as result of the ASRC without input attenuation.

Unfortunately I do not have many "normal" DACs these days (they are all pretty much audio interfaces / DSP + DAC) but the MOTU Ultralite Mk5 on my desktop cleanly passes the +3 dBFS 11.025 kHz via optical input to analog output if I attenuate 3 dB via the DAC volume control. So you do not necessarily need to attenuate in upstream software if you have no SRC, the DAC volume control works.

It is a bit disappointing that the miniDSPs do not have any flexibility to attenuate prior to the ASRC. Using a miniDSP with SPDIF input seems like one of the use cases most susceptible to intersample overs.

If I have understood this correct, does not this test show that interpolating upstream like CamillaDSP does it, is indeed a very good way of doing it instead of the dac? To quote @mdsimon2:

To me even more reason to use CamillaDSP as you can capture with something like the UR23 and then drop the level by a few dB in CamillaDSP to safely handle intersample overs.

This problem is new to me, so please bear with me if I have misunderstood something.

My example showed that you can capture a 11.025 kHz tone at +3 dBFS via TOSLINK to USB card, route that to CamillaDSP and attenuate as needed to avoid clipping if you want to upsample in CamillaDSP. You can than route that attenuated / upsampled output to a USB DAC. You can do this attenuation using a permanent gain reduction or using the CamillaDSP volume control. The other nice thing about CamillaDSP is that it will show if you are getting clipped samples, so you do not need any additional measurement equipment to determine if you are experiencing intersample overs.

Michael
 
My example showed that you can capture a 11.025 kHz tone at +3 dBFS via TOSLINK to USB card, route that to CamillaDSP and attenuate as needed to avoid clipping if you want to upsample in CamillaDSP. You can than route that attenuated / upsampled output to a USB DAC. You can do this attenuation using a permanent gain reduction or using the CamillaDSP volume control. The other nice thing about CamillaDSP is that it will show if you are getting clipped samples, so you do not need any additional measurement equipment to determine if you are experiencing intersample overs.

Michael
And if you are already committed to minidsp (ie you want to use Dirac) you can always have a small SBC between your source and minidsp running camilladsp or whatever allows you to bring the signal down a couple of db. I find camilladsp is good as a hub also.
 
And for absolute proof of CamillaDSP handling attenuation / upsampling correctly here is a measurement of the +3 dBFS 11.025 kHz tone at 44.1 kHz sample rate in to the UR23 TOSLINK to USB card, UR23 is the CamillaDSP capture device running at 44.1 kHz, 3 dB attenuation and upsampling to 96 kHz done in CamillaDSP, CamillaDSP playback device is a MOTU Ultralite Mk5 running at 96 kHz. Measurement below shows the analog output of the MOTU with no DAC attenuation in to a Cosmos ADC.

1677685988546.png


65536-point spectrum using Blackman-Harris 7 window and 8 averages
Input RMS -2.16 dBFS
-7.3 dBFS C, -5.4 dBFS A
-2.2 dBFS 22 - 22k UNW
-104.8 dBFS >22k
Distortion at 11,024.8 Hz, -2.2 dBFS:
THD: N/A based on 0 harmonics [20..22000 Hz]
HHD: N/A [10 .. 9]
N: -115.7 dB [20..22000 Hz]
N+D: -119.8 dBFS A
THD+N: -115.7 dB [20..22000 Hz]

Michael
 
Last edited:
Unfortunately I do not have many "normal" DACs these days (they are all pretty much audio interfaces / DSP + DAC) but the MOTU Ultralite Mk5 on my desktop cleanly passes the +3 dBFS 11.025 kHz via optical input to analog output if I attenuate 3 dB via the DAC volume control. So you do not necessarily need to attenuate in upstream software if you have no SRC, the DAC volume control works.

Michael

We can make an educated guess which dacs would most likely have it.

Given that the manufacturer would need to have knowledge of this issue and the issue arises from an error in recording, imo dacs from manufacturers with a sound production background are the most likely to have this feature.

To be identified as an issue the manufacturer would need to have sota testing in place. This rules out all dac manufacturers with measurements that are not very clean. Given what we've seen in test results, this rules out all big brand dacs.

As the penalty would be lower Sinad, we can also rule out sota dacs that compete for highest sinad.

This imo pretty much leaves only a few candidates that could have it. Rme being one imo.
 
Last edited:
We can make an educated guess which dacs would most likely have it.

Given that the manufacturer would need to have knowledge of this issue and the issue arises from an error in recording, imo dacs from manufacturers with a sound production background are the most likely to have this feature.

When you say this feature are you saying the ability to use the DAC volume control to attenuate and eliminate clipping due to intersample overs? Or DACs that have permanent attenuation to avoid intersample over clipping?

My MOTU definitely clips with the +3 dBFS 11.025 kHz tone if no attenuation is applied. However having digital volume control and the ability to eliminate this issue via digital volume control is a perfectly acceptable solution IMO.

To be identified as an issue the manufacturer would need to have sota testing in place. This rules out all dac manufacturers with measurements that are not very clean. Given what we've seen in test results, this rules out
all big brand dacs.

A slightly disagree with this as it is very easy to find, the spectrum is not exactly subtle and does not require a high resolution ADC to discover.

As the penalty would be lower Sinad, we can also rule out sota dacs that compete for highest sinad.

For a DAC with fixed level output I agree. Although I keep coming back to the concept of digital volume control eliminating this issue.

I am interested if anyone knows of a DAC with digital volume control that is still susceptible to intersample over clipping if appropriate attenuation is applied in the DAC volume control? I imagine there is something out there that has upfront SRC that suffers with this but it would be a bit esoteric.

I imagine most DACs are perfectly fine if they have digital volume control. Has anyone seen evidence to the contrary?

Michael
 
When you say this feature are you saying the ability to use the DAC volume control to attenuate and eliminate clipping due to intersample overs? Or DACs that have permanent attenuation to avoid intersample over clipping?

I meant permanent attenuation. But must admit I was not aware that it could also be fixed by digital volume control.
 
Back
Top Bottom