Lavry Engineering Paper on Hi-Res

[Miska]

The point is not that conversion at sample frequencies higher than 96kHz is “not accurate enough for audio;” it is that conversion at sample frequencies higher than 96kHz will always be less accurate than conversion at 96 kHz (or lower) with the same technology.

 

This applies only to multi-bit PCM. But with delta-sigma, higher the sampling rate, better the accuracy in audio band. One-bit conversion at 24 MHz rates can already give extremely good audio-band linearity with extremely simple and accurate circuitry.

 

Even with multi-bit PCM, using higher sampling rate and noise shaping combined with suitable analog filtering can improve accuracy significantly by reducing and shaping the LSB mis-match errors.

 

 

[Miska]

On the 96k vs. 192k sampling rate, this might be of interest to further muddy the waters

 

I wonder why they use 10+ year old DAC chip (AD1853) instead of more modern but still old AD1955, which would already perform better at 192k than at 96k...

 

 

[Miska]

For those interested in his opinion on DSD, there are a number of Posts on the Lavry Forum regarding this matter.

 

Funny part in those messages is that multibit PCM converters are practically dead. Much more than DSD. All the best performing converters are SDMs these days. And now we can do SDM in computers too instead of DAC chips, and with much better quality and SNR.

 

So the way to get best results today is to keep the data SDM all the way from ADC to DAC without going through two conversion steps to first convert it to PCM in ADC and then back to SDM in DAC. Especially because SDM is more space efficient format too and without the brickwall filter implications of the conversion.

 

 

[Miska]

Nyquist/Shannon theorem does not say a sampled file contains all the information of the original source

 

There are two often ignored aspects of the theorem vs. it's application to real world:

1) It assumes infinitely steep perfect filters - in real world "infinite" and "perfect" don't exist

2) It also assumes perfect sample timing and infinite sample accuracy

 

It also ignores transient and signal-change related parts, things like Gibbs phenomenon.

 

Also multi-bit ladder PCM converters can be significantly improved by noise shaping. This way it is possible to achieve more resolution in the audio band by utilizing higher sampling rates.

 

Just look at last and best multi-bit ladder DAC, PCM1704 and how it's performance improves by oversampling and noise shaping the input and using 768 kHz sampling rate and 24-bit resolution.

 

SDM on the other hand starts from the assumption, that since real world cannot be perfect and very accurate (due to manufacturing tolerances, temperature changes and such) - don't even try, but use that to your advantage.

 

Already in the past, radios became software defined (see http://en.wikipedia.org/wiki/Software-defined_radio ), now we can have software defined audio converters! That's my personal area of work.

 

 

[Miska]

Is it linear? Logarithmic? Some other curve?

 

It's a curve that can be modified by using noise shaping, so you can retain the low frequency resolution while increasing sampling rate.

 

But in the end, all the PCM bit depths of modern converters are fake, they just define how many bits the built-in DSP uses for external PCM communication.

 

When it comes to converters, my signal analyzer has 16-bit resolution at 10 MHz sampling rate.

 

 

[Miska]

Can you have 24-bit accuracy at 192K or not?

 

Looking at some recent "32-bit" DACs, it is possible to get fairly close to 24-bit accuracy.

 

But another thing is that traditionally with PCM it is thought that the accuracy would be flat within the entire frequency band (0-96 kHz for 192k sampling rate). But I find it more beneficial to focus the accuracy around 20 kHz audio band even at 192k sampling rate. So what I do is to trade about two to three bits in >20 kHz range to gain extra six bits in < 20 kHz range. But this is only possible at high enough rates.

 

In the end, it is simply about statistics, instead of having two samples for 20 kHz tone you have more than eight. If every sample has some percentage of random error, having more samples increases accuracy because the random error part cancels out.

 

 

[Miska]

I'd be curious, Miska, about what the calculations are that you did to determine 8x sample rates are reasonably sufficient.

 

I base my calculation on a filter that would have proper 20 kHz passband and enough attenuation. And the impulse response would fit into length of a half-wave 20 kHz sine.

 

 

[Miska]

Currently one of the best audio ADC is ESS9102 (http://www.esstech.com/index.php?p=products_ADC). It outputs 32-bit PCM up to 384 kHz sampling rates, but the actual converter has much less bits and much higher sampling rate. Probably similar to their DACs, meaning 6-bit at 40 MHz. But with noise shaping it achieves good audio resolution, meaning that the noise floor is heavily tilted towards high frequencies.

 

As you can see from the specs, it's SNR is worth 21 bits and linearity is worth 20 bits.

 

 

[Miska]

*Because* I can do all ahead in PC software and the DAC does nothing, I can easily compare any filtering means (useful for PCM).

 

I'm just doing the same for both SDM and PCM. I don't actually like to position myself for SDM any more than for PCM. But I do consider cost and reduced complexity of SDM to be beneficial.

 

Even though I don't mandate it, there are ways to do "NOS" DAC for either PCM or SDM. The biggest difference to Peter's approach is that I don't like those being filterless, I consider analog reconstruction filter being crucial part of the design. Discrete component SDM is also much more practically feasible.