Upsampling 16/44.1 collection a good idea?

[KSTR]

Is my understanding correct?

Corrent on all points

 

[antcollinet]

My understanding is that the primary benefit of oversampling is to improve the effective SNR with respect to quantization noise. Specifically, dithering is used to redistribute the quantization noise over a wider frequency spectrum extending well above the audible frequency range, and then the higher inaudible frequencies are filtered out. The audible difference primarily is detectable in quiet passages where the music signal is low, for example a decaying piano tone. See, e.g., https://science-of-sound.net/2016/01/quantization-noise-and-bit-depth/

But that only works at the adc stage, or when downsampling. It is not (if I understand correctly) possible to shape the quantisation noise in an existing quantised signal by upsampling. It is already baked in.

 

[danadam]

My understanding is that the primary benefit of oversampling is to improve the effective SNR with respect to quantization noise. Specifically, dithering is used to redistribute the quantization noise over a wider frequency spectrum extending well above the audible frequency range, and then the higher inaudible frequencies are filtered out.

AFAIK that's in AD process, not DA.

 

[CapMan]

want to keep every step in the chain working as perfectly as possible, so that at the end of the chain, all these "barely perceptible" and "questionable if perceptible" effects are minimized to the maximum degree, so that when summed their total is as close to imperceptible as possible.

Interested to hear what you do you do about your mood, mental state, hearing, external noise etc which all probably have more impact than any of these other refinements…just sayin’

 

[terryforsythe]

But that only works at the adc stage, or when downsampling. It is not (if I understand correctly) possible to shape the quantisation noise in an existing quantised signal by upsampling. It is already baked in.

AFAIK that's in AD process, not DA.

That is contrary to my understanding. See, e.g., Figure 4.3 in the following paper and the associated description: https://www.analog.com/media/en/tra...handbooks/Technical-Tutorial-DDS/Section4.pdf

See, also, https://www.analog.com/media/en/training-seminars/tutorials/mt-017.pdf

 

[antcollinet]

That is contrary to my understanding. See, e.g., Figure 4.3 in the following paper and the associated description: https://www.analog.com/media/en/tra...handbooks/Technical-Tutorial-DDS/Section4.pdf

That is talking about oversampling not upsampling. It is an adc process and must be done at the sampling time. Once the quantisation noise is all “in band” as it is when sampling at the nyquist frequency it cant be subsequently moved.

 

[terryforsythe]

That is talking about oversampling not upsampling. It is an adc process and must be done at the sampling time. Once the quantisation noise is all “in band” as it is when sampling at the nyquist frequency it cant be subsequently moved.

I respectfully disagree. See the following presentation titled "Digital Signal Processing Oversampled Ditgital to Analog Conversion with Noise Shaping": https://spinlab.wpi.edu/courses/ece503_2014/5-6oversampled_dac_shaped.pdf

From p. 2 of that presentation: "To minimize the effects of quantization noise (and to make the reconstruction filter easier to realize) we can oversample and shape the quantization noise just as we did with analog to digital conversion."

Also, I edited the post with another reference titled "Oversampling Interpolating DACs": https://www.analog.com/media/en/training-seminars/tutorials/mt-017.pdf

From the p. 4 of that paper: "Sigma-delta DACs operate very similarly to sigma-delta ADCs, however in a sigma-delta DAC, the noise shaping function is accomplished with a digital modulator rather than an analog one." Also, take a look at the paper's summary.

 

[tmtomh]

Why is this even still a discussion? Every modern DAC oversamples whatever you put into it, unless it's specifically designed not to (or to have a switchable option not to). The effective SNR of 16-bit content with noise-shaping dither is 120dB. And sample rate has nothing to do with quantization noise.

If anyone can provide proper evidence that converting 16/44.1 files to 24/88.2 files, or playing 16/44.1 files through a software app that converts them on the fly to 24/88.2, produces an audibly different noise level through any typical/modern oversampling DAC compared to just feeding the DAC directly with the 16/44.1 file source, please supply it.

Otherwise there's no need for this thread to continue.

 

[terryforsythe]

Why is this even still a discussion?

Because it is fun. I am enjoying it, and I suspect others are too.

In addition, some of us are learning from these discussions.

Otherwise there's no need for this thread to continue.

You are not required to read it. If you don't like it, just pass it by.

 

[terryforsythe]

For me, I am doing it seriously as I want to learn.

Prior to getting involved with this thread, I knew oversampling was used, but had never investigated why. This thread spurred me to do more research. Between results of that research and some of the posts herein, I have learned a lot.

 

[terryforsythe]

I love the beauty of on-line discussion.

I do too, so long as it is kept civil and respectful.

 

[antcollinet]

I respectfully disagree. See the following presentation titled "Digital Signal Processing Oversampled Ditgital to Analog Conversion with Noise Shaping": https://spinlab.wpi.edu/courses/ece503_2014/5-6oversampled_dac_shaped.pdf

From p. 2 of that presentation: "To minimize the effects of quantization noise (and to make the reconstruction filter easier to realize) we can oversample and shape the quantization noise just as we did with analog to digital conversion."

Also, I edited the post with another reference titled "Oversampling Interpolating DACs": https://www.analog.com/media/en/training-seminars/tutorials/mt-017.pdf

From the p. 4 of that paper: "Sigma-delta DACs operate very similarly to sigma-delta ADCs, however in a sigma-delta DAC, the noise shaping function is accomplished with a digital modulator rather than an analog one." Also, take a look at the paper's summary.

That is shaping the noise from the downsampling from (eg) floating point to 16 bit after the dsp. Note it says just as we did diuring adc. The q noise from the original adc is still there. The only noise being shaped is the second quantisation noise from downsampling to 16 bit After the floating point dsp.

 

[terryforsythe]

That is shaping the noise from the downsampling from (eg) floating point to 16 bit after the dsp. Note it says just as we did diuring adc. The q noise from the original adc is still there. The only noise being shaped is the second quantisation noise from downsampling to 16 bit After the floating point dsp.

Please point to which reference, and the passage of the reference, to which you are referring.

 

[terryforsythe]

That is shaping the noise from the downsampling from (eg) floating point to 16 bit after the dsp. Note it says just as we did diuring adc. The q noise from the original adc is still there. The only noise being shaped is the second quantisation noise from downsampling to 16 bit After the floating point dsp.

Also, see p. 2 of https://www.analog.com/media/en/training-seminars/tutorials/mt-017.pdf, which states:

OVESAMPLING INTERPOLATING DACS

The basic concept of an oversampling/interpolating DAC is shown in Figure 2. The N-bit words of input data are received at a rate of fc. The digital interpolation filter is clocked at an oversampling frequency of Kfc, and inserts the extra data points. The effects on the output frequency spectrum are shown in Figure 2. In the Nyquist case (A), the requirements on the analog anti-imaging filter can be quite severe. By oversampling and interpolating, the requirements on the filter are greatly relaxed as shown in (B). Also, since the quantization noise is spread over a wider region with respect to the original signal bandwidth, an improvement in the signal-to-noise ratio is also achieved. By doubling the original sampling rate (K = 2), an improvement of 3 dB is obtained, and by making K = 4, an improvement of 6 dB is obtained... (emphasis added).

 

[antcollinet]

Please point to which reference, and the passage of the reference, to which you are referring.

your first link. First section:

Often the precision of the signal processing algorithm is higher than the precision of the DAC, e.g., floating point processing and a 16-bit DAC.
…


In these cases, we have to quantize the algorithm output before sending it to the DAC.

So for example - a signal has been digitised, sent to some DSP which then processes it in some high resolution format, but only has a 16 bit dac to deal with, so must be re-quantised to 16 bit before conversion to analogue.

Think about it. If the quantisation noise of the original digitisation were to be subsequently shaped, it would have to be possible to somehow separate it from the signal. This however is impossible.

In the original sampling the actual analogue value is (say) between ADC values 500 and 501. It is closer to 501, so 501 is chosen. The quantisation noise for this sample is the difference between the actual analogue value and 501.

Now though the data is 501. It is impossible for any future processing to know what that noise is - or even if there *is* any noise. - there is no way to tell if the actual analogue value was exactly 501 or if not, where in the range 500.5 to 501.49999 it was. The noise is now permanently in the signal and cannot be separated. The noise is indistinguishable from the signal.

 

[antcollinet]

Also, see p. 2 of https://www.analog.com/media/en/training-seminars/tutorials/mt-017.pdf, which states:

OVESAMPLING INTERPOLATING DACS

The basic concept of an oversampling/interpolating DAC is shown in Figure 2. The N-bit words of input data are received at a rate of fc. The digital interpolation filter is clocked at an oversampling frequency of Kfc, and inserts the extra data points. The effects on the output frequency spectrum are shown in Figure 2. In the Nyquist case (A), the requirements on the analog anti-imaging filter can be quite severe. By oversampling and interpolating, the requirements on the filter are greatly relaxed as shown in (B). Also, since the quantization noise is spread over a wider region with respect to the original signal bandwidth, an improvement in the signal-to-noise ratio is also achieved. By doubling the original sampling rate (K = 2), an improvement of 3 dB is obtained, and by making K = 4, an improvement of 6 dB is obtained... (emphasis added).

And to answer that section.

The DAC is doing DSP. The upsampling interpolation is DSP, and the digital filter is also DSP. Whatever resolution that DSP is done at results in quantisation noise. (The value that would occour as a result of the calcualtion cannot be exactly represented within the resolution of the data words - so each time there is a quantisation error. The oversampling means this noise is spread over the wider bandwidth.

This does not alter the quantisation noise already embedded in the input signal.

 

[tmtomh]

A modern DAC can oversample internally but we are not sure that it is applying the best algorithim in doing the oversampling.

With NOS mode, user can select the oversampling algorithm that they want.

It’s not about best - it’s about sufficient to produce a result that’s equal to or better than human hearing ability.

 

[tmtomh]

Because it is fun. I am enjoying it, and I suspect others are too.

In addition, some of us are learning from these discussions.


You are not required to read it. If you don't like it, just pass it by.

Glad you are enjoying and learning. But threads are usually also read by people who don’t post in them, both now and in the future. So it can helpful - and imho very important - to distinguish clearly between an interesting theoretical, physical, or electrical question on the one hand, and an actual problem or issue on the other hand. And as @sunjam ’s reply to me illustrates, that distinction isn’t being made here.

 

[terryforsythe]

your first link. First section:



So for example - a signal has been digitised, sent to some DSP which then processes it in some high resolution format, but only has a 16 bit dac to deal with, so must be re-quantised to 16 bit before conversion to analogue.

Think about it. If the quantisation noise of the original digitisation were to be subsequently shaped, it would have to be possible to somehow separate it from the signal. This however is impossible.

In the original sampling the actual analogue value is (say) between ADC values 500 and 501. It is closer to 501, so 501 is chosen. The quantisation noise for this sample is the difference between the actual analogue value and 501.

Now though the data is 501. It is impossible for any future processing to know what that noise is - or even if there *is* any noise. - there is no way to tell if the actual analogue value was exactly 501 or if not, where in the range 500.5 to 501.49999 it was. The noise is now permanently in the signal and cannot be separated. The noise is indistinguishable from the signal.

Your understanding of the references is different than mine. Again, see p. 2 of https://www.analog.com/media/en/training-seminars/tutorials/mt-017.pdf.

See, also, https://en.wikipedia.org/wiki/Delta-sigma_modulation, which states:

Delta-sigma (ΔΣ; or sigma-delta, ΣΔ) modulation is an oversampling method for encoding signals into low bit depth digital signals at a very high sample-frequency as part of the process of delta-sigma analog-to-digital converters (ADCs) and digital-to-analog converters (DACs). Delta-sigma modulation achieves high quality by utilizing a negative feedback loop during quantization to the lower bit depth that continuously corrects quantization errors and moves quantization noise to higher frequencies well above the original signal's bandwidth...
...

Advantages of oversampling[edit]​

Oversampling converters instead produce a lower-bitdepth result at a much higher sampling frequency. This can achieve comparable quality by taking advantage of:

• Higher accuracy in time (afforded by high-speed digital circuits and highly accurate clocks).
• Higher linearity afforded by low-bit ADCs and DACs (for instance, a 1-bit DAC that only outputs two values of a precise high voltage and a precise low voltage is perfectly linear, in principle).
• Noise shaping: moving noise to higher frequencies above the signal of interest, so they can be easily removed with low-pass filtering.
• Reduced steepness requirement for the analog low-pass anti-aliasing filters. High-order filters with a flat passband cost more to make in the analog domain than in the digital domain.

Also, see the dithering and D/A sections of: https://classes.engr.oregonstate.edu/eecs/spring2021/ece627/Lecture Notes/chap14-Oversampling data converters.pdf

 

[CapMan]

believe scientists want to find the best solution available instead of just a workable solution. Isn't it?

Audio equipment is designed by engineers and made within the constraints set out by accountants. Perfect is the enemy of good enough !