PGGB is an expensive off-line upsampler that promises an extremely high resolution reconstruction of a lower-res recordings (like Redbook), using up to 256-bit floating point computation and avoiding the need to use sinc interpolation (according to the author).
If you read the often glowing reports from customers, it makes an amazing audible difference, despite taking some very serious PC hardware and significant time to upsample existing files, often taking hours or even days to complete and requiring 4x or more times the storage of the original files. Maybe it's really worth it?
Archimago did an evaluation of PGGB previously, and used my DeltaWave software to do the null analysis among other tests. I decided to see if I can reproduce his results and provide a little more details on what PGGB can or can't do. To do this, I picked an 44.1k album to upsample and left PGGB running overnight on an 8 core, 3.6GHz I9 64GB RAM computer. In about 7 hours, I only got 3 of the 10 tracks converted, but that was enough for some initial tests. These were upsampled to 705.6k with all the default/automatic settings.
My first test was to compare the results produced using PGGB and that produced by DeltaWave built-in upsampler. DW uses standard upsampling algorithms and 64-bit floating point compared to the 256-bit of PGGB. Both upsampled the same track to 705.6k. PGGB file was used as a reference, meaning that it wasn't modified during DW processing, only the comparison file is adjusted to match gain and phase of the reference file in DW. Assuming PGGB does something amazing, we should see clear differences between the DeltaWave upsampler result and that done by PGGB.
Spectrum plot... shows no obvious difference. The RMS level of the error after subtracting the two files (in time domain) is -187dB! A-weighted, it's down to -204dB:
Any differences when we subtract the two spectra? Nope. Looks like they are the same:
Phase differences? Doesn't get more a straight-line than this:
Let's see what the actual difference waveform looks like (again, this is computed in the time domain). Notice that even the highest peak value is below the level of thermal noise and overall, seems to be related to dither or quantization error at an extremely low level:
Here's the spectrum of the actual difference (null) file:
PK Metric audibility plot is just as you'd expect from all of the above:
Also notice the numbers across the bottom:
If you read the often glowing reports from customers, it makes an amazing audible difference, despite taking some very serious PC hardware and significant time to upsample existing files, often taking hours or even days to complete and requiring 4x or more times the storage of the original files. Maybe it's really worth it?
Archimago did an evaluation of PGGB previously, and used my DeltaWave software to do the null analysis among other tests. I decided to see if I can reproduce his results and provide a little more details on what PGGB can or can't do. To do this, I picked an 44.1k album to upsample and left PGGB running overnight on an 8 core, 3.6GHz I9 64GB RAM computer. In about 7 hours, I only got 3 of the 10 tracks converted, but that was enough for some initial tests. These were upsampled to 705.6k with all the default/automatic settings.
My first test was to compare the results produced using PGGB and that produced by DeltaWave built-in upsampler. DW uses standard upsampling algorithms and 64-bit floating point compared to the 256-bit of PGGB. Both upsampled the same track to 705.6k. PGGB file was used as a reference, meaning that it wasn't modified during DW processing, only the comparison file is adjusted to match gain and phase of the reference file in DW. Assuming PGGB does something amazing, we should see clear differences between the DeltaWave upsampler result and that done by PGGB.
Spectrum plot... shows no obvious difference. The RMS level of the error after subtracting the two files (in time domain) is -187dB! A-weighted, it's down to -204dB:
Any differences when we subtract the two spectra? Nope. Looks like they are the same:
Phase differences? Doesn't get more a straight-line than this:
Let's see what the actual difference waveform looks like (again, this is computed in the time domain). Notice that even the highest peak value is below the level of thermal noise and overall, seems to be related to dither or quantization error at an extremely low level:
Here's the spectrum of the actual difference (null) file:
PK Metric audibility plot is just as you'd expect from all of the above:
Also notice the numbers across the bottom:
- 85% of the samples in the two upsampled files match perfectly when expressed as 24 bit samples. Note that even simple random TPDF dither at 24 bits would result in only around a 50% match
- PGGB reduced gain by 0.1565dB, supposedly to eliminate intersample overs. DeltaWave adjusted the comparison track to match
- Again, the RMS of the difference is -187dB, A-weighted to -204dB...
