Convert 16-bit session to 24-bit session?

[lwilliam]

My understanding is that when converting from 16 to 24 bits, it simply pads the 16-bit sample byte with 8 zeros. By itself, that won't do anything and produces no sound improvement, but once you start making calculations on the data, then it's similar to only having 2 decimal places to get mathematical precision vs 4 decimal places: the information will be more accurate, and therefore will sound better.

Anything that re-calculates the waveform data would benefit from those 8 extra bits: EQ, Compression, even a volume change. Certainly summing will improve accuracy. As an example, adding 2.41 plus 2.62 will produce a certain level of accuracy, but adding 2.4121 plus 2.4179 will give greater detail. This isn't exactly analogous since that calculation method applies to fixed point math (protools HD), not floating point math (protools LE). Floating point math will produce more summing precision even with the first two since the decimal point literally floats based on the actual sum (less rounding). This doesn't mean that LE summation is better - as there are other factors involved. However virtually all native systems use floating point math for summing and PT HD uses fixed point math.

Dithering is not related. It's simply noise added to cover/disquise the quietest noise level in the audio file. It will allow for a cleaner removal (truncation) of those last 8 bits.

[daeron80]

That's right. But PT sums at the same bit depth whether your session depth is 16 or 24, so if a 16-bit session were completed except for mixing, and all mixing was to be done in real time, then there would be nothing gained (no pun intended) by converting the session to 24-bit. But there could be some gain by up sampling, depending on the plug-ins used.

It's true that EQ technically benefits from a greater bit depth. I haven't found it to be as audible as the change to a higher sample rate, in most cases that I've experienced. I'm sure it's a big help when doing multiple consecutive EQ processes, since the rounding errors would be a lot smaller. But what really helps EQ is spreading the errors over a wider frequency spectrum, especially pushing half of them up into the supersonic range.

[lwilliam]

Originally Posted by daeron80

That's right. But PT sums at the same bit depth whether your session depth is 16 or 24, so if a 16-bit session were completed except for mixing, and all mixing was to be done in real time, then there would be nothing gained (no pun intended) by converting the session to 24-bit. But there could be some gain by up sampling, depending on the plug-ins used.

Interesting point (PTLE does always sum at 32-bit floating point regardless of the session bit-depth). I also hadn't considered the benefits of any plugin upsampling (not sure how you find out if a particular plugin does it).

However, isn't summing done by reducing the gain for a channel times the number of channels, and then calculating the new sample data total so that it maxes out at 24 "ones" for the peak?

If that's true, then calculating the new reduced gain level for a 16-bit session would still involve those extra 8 bits on the master bus for rounding at least, no?

I would think also that reducing the gain of a 16-bit track in order to sum it would produce more precision by using those extra 8 bits simply for the gain reduction calculation.

I really don't know the algorithms used in this process, but I think we agree that there are distinct and often audible benefits to converting a session to 24-bits at just about any stage.

[this is sounding like a Gearslutz thread now.... ]

[flommer]

However, isn't summing done by reducing the gain for a channel times the number of channels, and then calculating the new sample data total so that it maxes out at 24 "ones" for the peak?

I think this part is backwards... The numbers are summed by dumping the full 24bit values into a 32 bit calculation and then that result is "down-sampled" or whatever the correct term is back to an equivalent 24 bit value.

If you reduced the gain to all the individual inputs first, you are are reducing the dynamic range across the board. The whole point of the 32 bit summing bus is to accommodate the larger values needed during the calculation.. I think that's why HD uses 40bit fixed point.. to have an even higher possible sum without "floating" away some bits on the bottom end (rounding)..

I may be way off on this, however...

[daeron80]

This page helps some:
http://www.digidesign.com/index.cfm?...0&itemid=15523

LE uses 32-bit floating point math (or did at the time that page was written, which is not noted). That provides tons of headroom and tailroom. Enough that, in practically any mix scenario, all available bits can be summed, and the result should be in bounds. There's so much headroom that, when opening a non-clipping LE mix in HD, it's actually possible to clip TDM plugins! The white paper on mixing in the box shows what the extra bits contribute; though it uses TDM for its examples, the same principle applies to LE, only with floating point instead of fixed math:
http://akmedia.digidesign.com/suppor..._Box_26689.pdf

Whether you're mixing a 16-bit LE session or a 24-bit one, summing is done using the same high res math. The plugins operate at the same high res. The bit stream passing through the stereo master bus is the same high res, too. Only the audio files themselves are 16-bit, and so only operations that end up storing their outputs to files in the session get truncated or dithered (depending on your prefs setting) down to 16-bit.

[daeron80]

If you want an experiential education in what happens at the bit level in LE, download the free VST "X-ISM" by SSL (if you have the VST-RTAS adapter). Import a CD track into a 16 bit 44.1 kHz session, a 24 bit 44.1 kHz session, and a 24 bit 48 kHz session. Have only the stereo audio track with the CD track on it, and a master fader with X-ISM on it. Play with the faders and watch what happens to the bits! Now start putting various plug-ins on the audio track one at a time, change the settings a little, as well as bypass each, and watch what happens to the 8 LSBs. Record a little noise from a pair of analog inputs (watching that on the bit scope will stop a purists heart), and do the same with it looped.

The funny thing is, I always thought that conversion to 24 bit added 8 ZEROS to the end. It adds 8 ONES. In PT LE, anyway. After watching the bits dance a while, it kinda makes sense that they should be ones. But I'd always been told zeros.

For those who don't have the adapter and don't have access to a bit scope, a summary:

1. In both 44.1 kHz sessions, the file retains 16 bit resolution as long as both faders are nul and no plug-ins (other than X-ISM).
2. In the 48 kHz session, the file has activity in all 24 bits, even when both faders are nul - sample rate conversion increases word length! I feel like I should have known that but it took be by surprise.
3. In all sessions, the slightest deflection of either fader produces activity in all 24 bits, as expected.
4. Regardless of the fader positions, adding 16 bit dither to the master (prior to X-ISM) returns the 8 LSBs to solid ones, just like the original file without gain change, as expected.
5. Plug-ins vary a lot in whether they:
a. screw with the signal even when bypassed
b. screw with the signal when active but all controls nul
c. treat the 16 bit stream as 24 bit, actually cleaning it up with small amounts of processing. And so on.
6. All 8 audio inputs of the 002R, whether mic or line, with all gain knobs at minimum, produce constant noise in the lowest 10 bits (of 24). Leaving only 14 bits totally clean. Ouch.

[kheftel]

Originally Posted by daeron80

6. All 8 audio inputs of the 002R, whether mic or line, with all gain knobs at minimum, produce constant noise in the lowest 10 bits (of 24). Leaving only 14 bits totally clean. Ouch.

I know this is an old thread, but would you mind telling me how many bits have noise in a 16-bit session? Are the lowest 10 bits still noise or is the number of noisy bits different?

[daeron80]

It would only be the lowest 2 bits if recording at 16-bit res. I can't imagine any situation these days where one would record analog at 16-bit, though.