In Search Of Accurate Sound Reproduction: The Final Word!

[mansr]

3 minutes ago, AJ Soundfield said:

You can't hear signal without transduction

Unless you have Bluetooth enabled cochlear implants.

[mansr]

3 minutes ago, semente said:

You may have heard about an experiment that Gradient Labs performed many years ago in which they compared the accuracy of several speakers by inserting a speaker into one channel.

The speaker would reproduce the signal/music in an anechoic chamber and this sound was then fed live into the corresponding speaker of the 2 channel setup placed in a listening room.

What was the outcome?

[mansr]

7 hours ago, CuteStudio said:

Disagreeing is a good, usually the source f the best discussions

 

For modern recordings - especially of rock/pop - mastering quality means a CD of Californication sounds significantly worse than the 192k lossy pre-master version. So the reality is we rarely get close to taking full advantage of redbook even.

 

That was my main 'dodgy', but it's interesting that you think redbook is OK. Digital source has been improved, most DVDs and SACD has a better matrix.

What is this matrix you speak of?

7 hours ago, CuteStudio said:

There are two major issues with redbook:

 

441.kHz. No one can build a real filter that works for that, unlike at 96 or 192 where the filter is trivial. So immediately we are into fudges to 'decompress' the waveform level-time matrix and various schemes are used for this. My favourite is a Behringer rate converter, which uses a Sharc DSP to calculate the missing points (the decompression phase) as accurately as possible and then the resultant 88.2k can be filtered with a HiFi filter. We still have an information limit of 22.05kHz which is low, but at least the filter can work. 

A rate converter is a digital low-pass filter. It is true that the sharp filter required for correct reproduction of 44.1 kHz material is more easily implemented digitally than with analogue components. That is why virtually all DACs upsample the input to 384 kHz or more.

7 hours ago, CuteStudio said:

A 96k source doesn't need a DSP to calculate the approximate missing points, because that's supplied in the information, because like it or not, at 44.1k part of what you are hearing is the guess of where the intermediate points go.

 

Which brings me to the terrible 16 bits. If 16 bits is Ok why is 8 bits bad? In SeeDeClip4 I can switch the mastering to 8bit and (at least in Chrome) I can listen to that, correctly dithered with a nice gaussian dither, and you know what? It sounds Ok. A lot better than you'd think. But not as good as 16bit, so if 16 bit is better than 8 bit, surely 24bit is better than 16 bit, especially as many real work DACs are now 18-20 bit.

Do you accept that human hearing has a limit? Do you agree that once enough bits are used to exceed this limit, there is no point in going further? It appears that somewhere between 16 and 20 bits is sufficient to capture anything a human could possibly hear. The audible difference from adding another bit becomes increasingly subtle with each one, so going from 16 to 24 bits will be barely noticeable while the difference between 8-bit and 16-bit resolution is readily apparent. In fact, even 16 bits is plenty for most music in normal listening conditions. You'd need a very, very quiet room to have any real benefit from more.

7 hours ago, CuteStudio said:

Many adherents of 16 bit then claim dither is the saviour, and point toward the maths that proves it's perfect. And the maths is right, for a 1kHz signal it pretty much IS perfect. But not the music. The maths is correct for a continuous wave, not for a short, transient one. If you study waveform shapes have a look at a quiet HF part and you'll see for say a soft cymbal strike there are shapes to the waveform. These shapes are wrong with a low bit + dither, and can only be correct with a higher bit rate. 

The maths you speak of is generic. It applies to all waveforms.

7 hours ago, CuteStudio said:

Linearity:

The 16 bit scale is 32767 bits per side of 0v, so you'd think that the resolution was 1/32767, but you'd be wrong for a reason that few people talk about: the logarithmic loudness of sound. What this means is that the distortion rises as the sound level falls. 6db down and you are at 15bit audio. That trailing ambience at -60dB? Welcome to 6bit audio.

 

Remember that 8bit music I said was quite good? Digital is 6dB per bit (each bit halves the signal) so 48dB down and most classical listeners will spend quite a bit of time oohing and aahing at basically 8bit digital music.

The magnitude of the distortion products from undithered quantisation depends on the bit depth, not on the amplitude of the signal. The distortion level at -60 dB signal is the same as full scale. The signal must of course be higher than the quantisation level. Anything lower is lost completely. With dither, even signals below the quantisation level are captured although at some point they are completely drowned out by the dither noise. In other words, the bit depth determines the usable dynamic range, and anything within this range is conveyed equally well. Dithered 8-bit actually has enough dynamic range that typical pop or rock music hardly suffers at all. With classical music, the dither noise becomes audible during relatively quiet parts.

7 hours ago, CuteStudio said:

As for speakers, it's simply that the distortion is higher than all other parts of the chain (poor mastering excepted), 10-20% distortion is not unusual, there are some very good speakers around, but generally if people want a better sound the best thing to change is the speakers IME.

Speakers are indeed the least accurate component in an audio playback system. There we agree.

 

[mansr]

18 minutes ago, CuteStudio said:

The XY matrix of time vs level. Imagine analog as a wave drawn on a whiteboard.

The digital matrix can then arrange dots in a grid that approximate that waveform. Then we store the dots, put them on a shiny disk and Fred plays that disk, which is where the dots are used to form an approximation of the analog waveform. If you imagine a 44.1 16 bits matrix as a grid of dots, then if you fill in twice as many columns (88.2kHz) you can see the grid is denser. Then add 255 levels between each row (24 bit), and you have a denser grid still.

This is why I use the term matrix, because what we are actually doing is putting a matrix over the waveform and choosing some squares/dots.

I see what you're getting at, but that's not an accurate view of what's actually going on.

18 minutes ago, CuteStudio said:

" It appears that somewhere between 16 and 20 bits is sufficient to capture anything a human could possibly hear. "

 

So what's the resistance to 24bit?

Why does it MATTER if 24bits is better than our hearing?

Is that really the game: to just up the tech enough to fool the average ear? Why can't we just standardize on something better? It's hardly difficult. We casually buy 1TB disks, 3GHz multi core processors and phones with 2GB of RAM and we're seriously arguing that we only need 16/44.1? Why are we doing this?

Why even the discussion?

We don't apply the 'just good enough' criteria to any other part of our Hifi so why apply it to the recording format?

Recordings should absolutely be done in 24-bit. There are a multitude of benefits to that. Distribution may or may not benefit from a resolution beyond 16 bits. Then again, absent any constraints of a physical medium (e.g. CD), there's no practical reason to not use the full 24 bits.

18 minutes ago, CuteStudio said:

"The maths you speak of is generic. It applies to all waveforms."

The maths applies to all continuous waveforms. Try it on a waveform 4 samples long. Not good. Worse than no dither. Dither is a way of averaging level errors over time: but on transient events you have no time, you want them accurate right then, not 200 samples later.

If we all listened to church organ music I'd agree that dither was a good answer, on transients it is demonstrably inadequate: there is no Free Lunch.

We are always dealing with band-limited signals. There are no discontinuous waveforms.

18 minutes ago, CuteStudio said:

"The distortion level at -60 dB signal is the same as full scale."

Mmm - I'm not getting that here: 

At 0dB you have 16 bits, with a quantisation distortion of 1/65535 = 0.0015%

At -60dB you have 6 bits, with a quantisation distortion of 1/63 = 1.59%

The absolute level is the same. If it is low enough to be inaudible, it doesn't matter what the signal level is.

[mansr]

57 minutes ago, CuteStudio said:

I Always though a 96dB filter in 0.05kHz was impossible but I'm willing to learn!

Where did you get that transition bandwidth? If the audible limit is 20 kHz, that gives you 2.05 kHz within which to achieve the necessary attenuation. This is easily done with digital filters.

[mansr]

19 minutes ago, CuteStudio said:

Yes you are right, 2.05kHz if one is aiming for 20kHz. 0.05kHz was a math-typo.

 

I said:

 

"44.1kHz. No one can build a real filter that works for that"

 

Referring to the DAC anti-aliasing filter which has to be analog.

As I said, just about all DACs upsample digitally to simplify the analogue filter.

[mansr]

19 minutes ago, CuteStudio said:

Surely upsampling involves digital interpolation (estimation)  of the intermediate values between the original data points? Is result then still redbook at all?

 

The process of mastering to 44.1 and then upsampling at the playback end could be described as a form of lossy compression because essentially you are throwing away data to create the 44.1 waveform,

This process is correctly termed band-limiting.

19 minutes ago, CuteStudio said:

and then estimating what those values may have been during the upsampling.

If the original recording had frequency content above 22.05 kHz, this is irretrievably lost, and upsampling does not involve guessing what it might have been. Rather, upsampling interpolates sample values such that the bandwidth is extended while preserving as closely as possible the frequencies in the input.

19 minutes ago, CuteStudio said:

The need to do this appears to corroborate my initial (oddly controversial) contention that anti aliasing filters don't work (very well) on the 44.1kHz redbook standard.

Digital filters work very well indeed. Analogue filters are trickier, but not impossible, which (I'm starting to feel like a parrot) is why DACs upsample digitally before applying a simpler analogue low-pass filter. Nothing is remotely controversial here.

[mansr]

2 hours ago, Ralf11 said:

I also sometimes hear that integer value upsampling should be used, but it is not clear ot me why that is advocated, tho it makes the FFT math easier.

There is no difference in accuracy between integer and non-integer ratios.

[mansr]

13 minutes ago, jabbr said:

Probably the DAC only needs one master clock and rates could be upsampled to a 2^n multiple of this (typically n is negative)

but DACs usually have two clocks, one for the 44100 hz family and the other for 48000. It's much easier for the DAC to generate the BCLK by division. 

Yes, it's easier to make a chip where all clocks have small integer ratios. That is why power of two multiples are common. If done in software, these constraints don't apply, and arbitrary ratios can be used.

[mansr]

9 hours ago, fas42 said:

The why's and wherefore's will be highly variable, every situation will be different - some years back I was idly fooling around with the playing of music files on a desktop machine, just using the inbuilt, low cost DAC on the motherboard, and simple side monitors. Tried upsampling; hello!! - quality is better! The higher the rate the better the quality got; I ended up at ridiculous levels, MHz sampling rates - and I always gained. Of course, the source music file was now staggeringly huge, Gigabytes for a single track!

 

So, a curiosity. Totally unusable as a real world solution, but, proved a point. Why it worked on that machine could be a whole subtle combination of factors; which would be irrelevant for the next setup.

I don't believe you. Cheap on-board audio invariably uses the Intel High Definition Audio (HDA) specification which supports only up to 192 kHz sample rate.

[mansr]

6 minutes ago, fas42 said:

Of course it doesn't make sense if looked at conventionally - which is why I stated, "Why it worked on that machine could be a whole subtle combination of factors". For me at the time those were the results I got - and since it was of no value for ongoing use I didn't bother going anywhere with it.

Why don't you tell us what hardware and software you used for this experiment? Perhaps we could learn something.

[mansr]

2 minutes ago, fas42 said:

Probably of little value for most - the machine is a "commercial" quality Compaq, very old now - running XP. The player software was an old, cut-down freebie version of Nero, which just happened to give the best SQ of all the software I tried - current foobar2000 was nowhere in the race. Zero DSP, all fiddling of the source as it passed through was reduced to a minimum.

And how did you create the super high sample rate files?

[mansr]

20 minutes ago, fas42 said:

Audacity. The latest versions use a converter as good as any, sourced from the SoX project. The software doesn't nominally do extreme sample rates, but if you feed in the correct numbers then the conversion is done. We're talking very, very long periods of time to do the upsampling, at the best quality - even saving the output files takes some time.

I suspect what actually happened was your high-rate files were downsampled for playback, either by the Nero player or by the Windows audio system. Whatever changes in sound occurred were artefacts of the up-then-downsampling process, not something to do with high sample rates per se.

[mansr]

10 minutes ago, esldude said:

I don't know if the Redbook actually spells that out.  One would think flat response to 20 khz and 96 db down at 22,050 hz.

I found a copy of the spec. It doesn't say anything about the D/A conversion.

[mansr]

4 hours ago, CuteStudio said:

So AJ, have you found the filter yet?

Just post up the schematic

Here's the full schematic of the Sony D-50 CD player. The lowpass filter is clearly marked in the top right corner.

[mansr]

1 hour ago, AJ Soundfield said:

Its not real, no one can build a 44.1k filter

And here's the measured frequency response of that CD player:

[mansr]

1 hour ago, SoundAndMotion said:

Many critics of the value of relaxed filters that higher sample rates allow, argue that whatever a steep filter does (phase is all I can think of, is there more?), it is inaudible. "Whatever" does seem to be audible (at least under the strict conditions you listed).

Passband ripple is another filter characteristic that could possibly be relevant. With a 20 kHz cutoff, it does extend into the audible region. Of course, for any reasonable filter, this ripple is far smaller than the variations from speakers and room effects.

[mansr]

23 minutes ago, CuteStudio said:

It's a complex filter, several RC sections, two active parts and a band-pass with 3 inductors: not one for the purists!

Nobody ever said it was simple. Quite the opposite, in fact.

23 minutes ago, CuteStudio said:

Interesting graph, if real! It appears to show a level reduction from -78.5dB to -105dB which appears to be a full 26.5dB cut, which is marvellous.

Of course it's real. I measured it myself. The player may be 35 years old, but it's still in perfect working order.

[mansr]

6 minutes ago, CuteStudio said:

So I noted it seems to have about a full 26.5dB cut, but if I look at the 22.05 position (as close as I can) it looks to be only about 13.25dB down at the point, rather than the required 96dB, can you confirm that the vertical scale is in dB?

Yes, it's dB.