Actually, 24bits is used for recording and mastering because it gives more dynamic headroom and makes recording and mastering easier, and for higher sampling rates many say it's not needed at all...
something to read Sampling theory
As pointed out, mastering is entirely different from playback: in the former plenty of headroom is needed because one can never be sure what the loudest sound will be, and it allows people to do crazy things in editing and get away with it. The final product, for playback, has a locked down, set in stone dynamic range - and for that 16 bits, intelligently used, are fine ...
Then you have never recorded material yourself. 192kHz is much better than 44.1kHz. 64-bit is also quite a lot better than 24-bit. 32-bit, only slightly better.
I thought 32bit floating point was the norm for mastering, but with 64bit becoming the norm, it seems reasonable to go to 64bit.
DSD at somthing like 5MHz seemed better than PCM, but the format cannot be used for mastering. I do wonder if anything is altered in the conversion process.
I think mentioning these formats are for the purpose of determining what to consider when conducting listening tests to try to minimize sound coloration upstream as much as possible.
DSD at somthing like 5MHz seemed better than PCM, but the format cannot be used for mastering. I do wonder if anything is altered in the conversion process.
I think mentioning these formats are for the purpose of determining what to consider when conducting listening tests to try to minimize sound coloration upstream as much as possible.
Very interesting, higher sample rate may not be better? I need to read it carefully.
There is a cool guitar amp simulator that uses 64-bit processing. Sounds cool, but nothing beats the real thing. I would prefer recording to tape and then convert it afterwards. But so much work.
Wow, I have been too lazy to find out how come I didn't like high sampling rate.
I could not understand what that was talking about. Many things are left unanswered in that report.
1. The measurement of accuracy is not scaled to the actual signal level with direct comparison between the sample rates.
2. The signal levels have no relation with actual music spectrum content. For example the closer you get to 20kHz, the lower the actual levels are. What would the error levels be when you use signal levels closer to real spectrum content?
Why would higher sampling frequency sound worse? Really depends on how hardware is designed which is really on a case by case basis. The design also effects USB cables used. Yes, in many cases digital cables so make a difference, and it seems primarily due to circuit ground layout and how it influences they analog section.
1. The measurement of accuracy is not scaled to the actual signal level with direct comparison between the sample rates.
2. The signal levels have no relation with actual music spectrum content. For example the closer you get to 20kHz, the lower the actual levels are. What would the error levels be when you use signal levels closer to real spectrum content?
Why would higher sampling frequency sound worse? Really depends on how hardware is designed which is really on a case by case basis. The design also effects USB cables used. Yes, in many cases digital cables so make a difference, and it seems primarily due to circuit ground layout and how it influences they analog section.
Sadly new tape seems to be a dieing market. I am not sure RMGI or ATR will still be making it in ten years. It is by far my favorite medium.
IMHO 20bit is enough for the ADC.
32 bit is enough for recording the data.
That allows for processing of the data.
But I do think 384k HZ.
Is best.
Easy rule, multiply by 6 ... so, 16 bits gives 96dB, 20 means 120dB, 24 equals 144dB, 32 is 192dB - and 64 scrapes in at 384dB - start of the universe type stuff ...
Easy rule, multiply by 6 ... so, 16 bits gives 96dB, 20 means 120dB, 24 equals 144dB, 32 is 192dB - and 64 scrapes in at 384dB - start of the universe type stuff ...
Thank you, but that was rhetorical question...
There are many excellent reasons to sample at 3-4 or more times the highest frequency of interest. From the real-world perspective it makes the inevitable analog low pass filtering MUCH easier. From the mathematical perspective you CAN theoretically record up to the Nyquist frequency, but in practice it is impossible without aliasing since infinite slope filters don't exist. As for playback, you still have to filter it, and if you want any semblance of phase coherence at high frequencies you had better be sampling at much more than twice the highest reproduced frequencies.
I touched on the sampling above, now I will touch on bit depth. 96dB dynamic range doesn't even match the S/N of the better amps and preamps available, let alone the human auditory system. Then there is the higher level of quantization error which is most noticeable at low levels.
The fact is, Redbook was designed to be a merely adequate consumer format. Bit depth and sampling rate were limited by the technology of the time - Sony and Philips wanted a disk of a certain size (small enough to foit in car stereos and portable players) that held a certain number of minutes of playback, and 16/44.1 was the best that could fit in the space available. Due to the limited feature (pit) size available at the time, 24/96 would've required an LP-size disk (or larger). Thus, we have been saddled with "Adequate Sound Forever" for over two decades.
24/384 is the norm for mastering. Floating point is relatively new, and offers some advantages and some disadvantages. A 32 bit floating point has a HUGE dynamic range - much larger than a 32 bit integer. However, it is not constant - the resolution is highest at numbers closest to zero, getting progressively coarser as get near the limits of the exponent. In practical terms it is plenty for audio - anyone worried about it could use double precision for any math being done. It's simply a matter of choosing the precision when you write the C code.
32 bit floating point does not exist because 24 bit is inadequate. It exists due to mathematical reasons. If you convert at 24 bit integer to a 32 bit integer and start performing multiple divisions you can quickly generate gross rounding errors. 32 bit floats give reasonably low rounding errors, and all modern cpus have hardware floating point division.
The reality is that 96dB is perfectly fine - one can do tests to 'prove' to oneself that this range in any normal listening environment is no problem at all, something I've done on a number of occasions and it always verifies that situation. Of course, one can do ridiculous experiments, grossly exaggerating gains in various ways which have nothing to do with how one listens to real, recorded music ...
A lot of this history is what I had remembered as well. The 24/384 sort of confirms my question about the sample rates. It always seemed to me that cutting any wave into at least 10 segments was more adequate. The level resolution vs time resolution simply are miles apart in the critical frequency range. In any event, digital content is going to be more accurate than the vinyl and tape content, we just need to get the electronics right.
Implementation issues. DACs aren't perfect devices - they introduce errors. The most significant errors for audio (as opposed to for instrumentation applications) turn out to be the dynamic ones. Glitches are generated every time a DAC's output code changes - R2R ladder converters are particularly poor for this. The following analog circuitry has to settle at each code change, which is in effect a step change. So does the analog circuit slew-rate limit? If so that's another error. Then there's settling time - only pure exponential settling introduces no error.
All the above issues get worse with increased density of code changes - i.e. higher sample rates.
Should one try to avoid all these problems of multibit converters and instead go sigma-delta architecture (as most consumer chips are these days) then you'll open a whole new can of worms as regards introducing noise modulation.