-4.2dB (70% volume) makes a dramatic difference at low bitrates for Lossy Encoders?

[Lynx_TWO]

I just stumbled on this phenomenon and wanted to get some feedback. I have been testing different audio codecs looking for one to duplicate my music library in ultra-small files at the lowest bitrate while still being somewhat pleasant to listen to. The lower the bitrate, the better, but I was shooting somewhere between 24-48kbps.

In this case, perceptional losslessness is not the goal rather, to have something with the least amount of nasty artifacts so I can put my entire library on a thumb-drive, and also just to do it.

The codecs I tried were as follows in dBpoweramp.

MP3 V9
MusePack -telephone
OggVorbis -q -0.1
FDK AAC encoder -m2 -a1 -p 29 (HEv2, afterburner on, results in around 43kbs)

The following I used foobar2000 to encode:

Nero AAC (using foobar2000 to encode)
qaac (Apple's AAC encoder, using foobar2000 to encode)

The FDK AAC encoder sounded the best at the lowest bitrate, even beating OPUS, at least to my ears.

What was interesting is, and this is probably dating myself, I remember the QDesign audio codec back from Quicktime 3 had always suggested lowering the input volume of audio for the encoder to 70%, or -4.2 dB. I did this using the Normalize DSP in dBpoweramp, decoded them back to FLAC and then analyzed the files. The results are quite interesting:


Normal input (no volume change) - Waveform Statistics of the FDK AAC HEv2 at -m2 -a1 -p 29



Normal input (no volume change) - Visual Waveform of the FDK AAC HEv2 at -m2 -a1 -p 29



-4.2dB input (70% of original Volume) - Waveform Statistics of the FDK AAC HEv2 at -m2 -a1 -p 29



-4.2dB input (70% of original Volume) - Visual Waveform of the FDK AAC HEv2 at -m2 -a1 -p 29



-4.2dB input (70% of original Volume) - Visual Waveform of the FDK AAC HEv2 at -m5 -a1 (Full Quality Low Complexity)


It seems there is a pretty clear advantage to lowering the input volume when encoding to low bit-rates, especially since lossy codecs are bit-depth independent (no dynamic range is lost). I did notice when doing this with LAME at V9, all the hard squeaky artifacts were gone after changing the input to 70%. Of course, it still sounded bad :D

It also appears there may even be an advantage to lowering the input a bit even when encoding at 256 kbps AAC, although not as dramatic.

Spoon, any thoughts on this? I wonder what is going on here and if whatever IS going on here could also be used to declip audio!?!

Thanks for all you hard work! Looking forward to purchasing dBpoweramp 16 Reference!

[Lynx_TWO]

Re: -4.2dB (70% volume) makes a dramatic difference at low bitrates for Lossy Encoder

I just tried this on Death Metallic, the LOUDEST CD on record. The input volume needs to be at -6.4dB, or 47% of original to avoid clipped samples, so I think that's where I'll leave the normalizer for encoding low bit-rates.


No volume change



dBpoweramp Normalize DSP set to lower volume by -6.4dB before encoding

[Spoon]

Re: -4.2dB (70% volume) makes a dramatic difference at low bitrates for Lossy Encoder

-4db is almost halving the signal, hence the reduction in bitrate (as there is less to encode).

[Lynx_TWO]

Re: -4.2dB (70% volume) makes a dramatic difference at low bitrates for Lossy Encoder

-4db is almost halving the signal, hence the reduction in bitrate (as there is less to encode).

Well, I think there is a lot more going on there... 16 bit is theoretically 96 dB dynamic range (really more than that with dithering) and most new music uses only about 10-12dB of the available range due to the aggressive dynamic compression to sound louder (but REALLY fatiguing on the ears), so really, a -4.2 or even -6.4dB change is only going to lower the dithered noise floor that much further with 16-bit audio. I suspect dithering noise is the only thing you are showing the encoder less of. Take a look at the waveforms as well. The encoder clearly has a de-clipping affect. The Dynamic Range tool shows Death Magnetic goes for a DR of 4 to a DR of 9 after lowering the input volume to -6.4dB before encoding. Pretty big difference. Interestingly enough, using a high bitrate and the Low Complexity model does not have as much of a declipping affect, probably due to maintaining a more faithful representation of the original signal, where the low bitrate version is much more of a synthetic re-creation?...

I suspect the real reason for the bitdepth reduction is inter-sample clipping tends to cause high frequency distortion, which is expensive to encode, but then I am just speculating... Still the affect is interesting, no?