Nord User Forum

Concert grand pianos have about a 50dB dynamic range. I think that would only be about 12 bits of range. And I doubt the audience hears the full range— the quietest sound heard by the player and the quietest sound heard by someone at the back of the hall are different so the player cannot use the entire range.

There is also the issue of resolution— how many fine graduations of dynamics can you distinguish in that 50 dB range, which might call for large quanta, but midi only represents 7 bits of velocity (0 to 127 and 2^7=128), so that’s all the resolution of dynamics you get with a digital piano. Intermediate levels can be interpolated, but that is already an approximation, and errors will limit the increase in resolution to 256 levels or 8 bits. This is probably why the midi standard was set to 7 bits as instrument samples for digital keyboards were 8 bits when the midi standard was set. I don’t think digital piano manufacturers want to build keyboard actions with 2^16 velocities and produce samples with 2^16 velocity layers, much less 24, so resolution within the dynamic range is not a driver for 24-bit quanta.

While human hearing has an audible range of 20 bits of dynamics, that includes ear-bleeding loud sound pressures like what would be heard by airport ground crew standing next to a jet if they did not wear ear protection devices. It is doubtful you want your released music to have a range from the faintest whisper to levels requiring ear protection, nor would most playback systems in use be able to handle that. So what is the use of having the full 20-bit dynamic range of human hearing encoded in our music? Answer: none.

I think the benefits of 24-bit samples in keyboards like Dexibell and Nord are in recording the instruments, not in playback. Imagine you are recording an acoustic piano. You record 24-bit quanta so you can set all the preamps and op-amps in the audio path during recording conservatively and not worry about a noise floor. You have extra bits of resolution so that the digital noise from post-processing (floating point roundoff errors) does not accumulate into the signal you care about. So you wait to dither down to 16 bits after all digital manipulations are completed, and release the recording in 16 bits.

But if you instead recorded with a digital piano, then if the piano uses 16 bit samples, the content is already 16 bits, which means it is as if you chose to dither the analog recording of the previous paragraph down to 16 bits before you did post-processing, losing the benefit of removing digital noise from post-processing even if you resample at 24 bits.

As an aside, there could have been one benefit of 48kHz sample rates for playback, which is that it would make it easier to implement the low pass filter in a DAC that filters out noise at the top end of the spectrum, the so-called brick wall filter in DACs. DAC chip manufacturers have nonetheless already implemented the filters that work well for 44.1kHz. These have to be very steep to filter down to effectively zero by 22kHz. A 48kHz sample rate would have enabled a less steep filter to be implemented (down to zero by 24kHz) , which is easier. Early CD players used lower frequencies for the low pass cutoff and early CDs were mastered to be bright to compensate. But this has not been an issue for the last 30 years or more.

Good rundown there.

And props for noting the dreaded "emphasis" of early consumer digital.

Thanks. I would add that for recording analog, especially acoustic instruments, midi resolution does not come into consideration, and there is an open question of size of quanta needed to capture the subtle nuances of dynamics. Can music played by an individual musician, ensemble, or orchestra have more than 2^16 dynamic gradations, and is the difference detectable by a listener if they are rounded to the nearest 1 in 2^16? I don’t know the answer, and I doubt it is fully clear cut, but I think most audio engineers consider dithered 16-bit digital audio to be sufficient not to detect improvements in larger quanta.

It is nonetheless interesting that Sony drove the Redbook standard of 16x44.1kHz PCM and also was the driver for marketing SACDs that use DSD encodings. D/A conversion for DSD is much easier than for PCM so there would have been some wisdom for using DSD as distribution medium. On the other hand SACD transports require greater precision than CD transports, observable by the high frequency of SACD players, especially cheaper ones, developing alignment problems and ceasing to recognize SACDs, so there would be cost trade offs for using DSD for discs.

We should also make a distinction between listening to raw samples and listening to a complex sample player/synth engine.
A 24 bit engine will sound better than a 16 bit engine, but both would not be the top of the quality.
For example, certains DAWS use a 64 bit floating point engine (ok, and they expecially need them when mixing and processing a lot of tracks).
A 24 bit sample give you more possibility in processing.
Why all this ? Many reasons, including computational errors (remember, digital processing *always* introduce errors), headroom in the final signal or within the gain chain, etc.

Maurizio

maurizio wrote:We should also make a distinction between listening to raw samples and listening to a complex sample player/synth engine.
A 24 bit engine will sound better than a 16 bit engine, but both would not be the top of the quality.
For example, certains DAWS use a 64 bit floating point engine (ok, and they expecially need them when mixing and processing a lot of tracks).

The normal hearing of a human below about age 40 can discern 20 bits of dynamic range. Once you are above that, you will not be able to distinguish differences. With dithering, 16-bit content is close enough to 20-bit dynamic range that most (but maybe not all) listeners would not be able to discern a quality difference.

The fact that some DAWs or other DSP implementations use 64-bit floating point is a performance issue, not a quality issue for 24-bit data. You only need 3 bits to represent the numbers needed to verify that 2+2=4, but you can do that calculation with 64-bit floating point if you want. The issue is avoiding overflow of a word. If you have 32-bit samples and 32-bit floating point, it will be inefficient to implement algorithms that do not have overflow errors because more than one 32-bit word will be needed to store a sample data point. But either 32 or 64-bit floating point arithmetic will work fine with 24-bit or 16-but samples.

A 24 bit sample give you more possibility in processing.
Why all this ? Many reasons, including computational errors (remember, digital processing *always* introduce errors), headroom in the final signal or within the gain chain, etc.

Maurizio

This was articulated above. If you want to release 24-bit content, work with 32 bits. But if you want to release 16-bit content, working with 24-bit samples is fine. But if the source is digital, you may have 16-bit samples to start with and rendering it as analog and up-sampling to 24-bits will not replace the info that was already lost in the original 16-bit sample.

A 64 bit floating point engine really is only equivalent to a 52 bit sample, if we're making a comparison.

(52 bits is plenty, since we're comparing it to 24 bits.)

Here's a good way to better understand what a 64 bit floating point number actually looks like (also 32 bit and 16 bit):

http://evanw.github.io/float-toy/

As I understand it and I may be wrong. With the Nord hardware especially latest equipment, sample requirements are 16 bit 44.1khz stereo? If this is any higher please let me know, I'm sure however that the stage 3, etc still rely on sample data going in to NSE as 44.1khz 16 bit stereo / mono waveform data..

lew

I only have the details from the youtube video Nord put out as a tutorial on the new Nord Sample Editor 3, but it says to bounce your sample at 44.1kHz and 16 bits, so yes, Lew, I think this is still the internal sample rate for the stage 3.

Edit to add: I'm guessing that intermediate calculations in the stage 3 engine are done at a higher precision. That's pretty normal, and ensures there is minimal loss of precision in the output after effects processing, volume, mixing of all outputs etc. It won't magically make a 44.1/16 sample have more precision, but it prevents any needless loss during intermediate calculations.

Edit #2: Looks like Lew deleted his account and all his posts. If you're reading this later, that's why it looks like I'm talking to a ghost. Lew, you are missed.