ReplayGain Implementation

[Carson Dyle]

Does Rockbox read and utilize ReplayGain peak fields, perhaps to avoid clipping?

REPLAYGAIN_TRACK_PEAK
REPLAYGAIN_ALBUM_PEAK

Or does it only use the gain values?

REPLAYGAIN_TRACK_GAIN
REPLAYGAIN_ALBUM_GAIN

What about reference loudness?

REPLAYGAIN_REFERENCE_LOUDNESS

[tpijag]

Gain type and whether to protect from clipping is up to you - via the settings.
Regarding reference loudness, I assume that is adjusted via pre-amp settings.

For the sansa clip that I use it is all in the manual. http://download.rockbox.org/daily/manual/rockbox-sansaclip/rockbox-buildch7.html#x10-1220007.7

[Carson Dyle]

Right, I'm aware of the available settings.  I was just wondering whether the other RG values are used in the implementation or if they were ignored.

[saratoga]

Right, I'm aware of the available settings.  I was just wondering whether the other RG values are used in the implementation or if they were ignored.

You can't prevent clipping without using the PEAK tag.  Literally whats done is that you take the peak tag and compare it to 1.  If its greater, you divide every sample in the song by that amount.  If you don't have that tag theres no way to avoid clipping (since you have no idea if the song is going to clip).

[Carson Dyle]

You can't prevent clipping without using the PEAK tag.  Literally whats done is that you take the peak tag and compare it to 1.

Thanks. Are you talking about the raw value of the PEAK fields or some computed value?  The RG track and album PEAK values can't be greater than 1.

[saratoga]

Are you talking about the raw value of the PEAK fields or some computed value?

The raw value.  The peak is exactly what it sounds like.  The amplitude of the largest sample relatively full scale (1.000000).  Hence its a little odd of you to ask if the peak information is used to determine the peak. 

  The RG track and album PEAK values can't be greater than 1.

Looking at my player, something like 80% of the songs on my player have peaks > 1.  Pretty much any modern music will come out well over 1 if MP3 compressed because lossy compression is only approximate and need not have anywhere near the same amplitude as the input signal. 

[Carson Dyle]

Looking at my player, something like 80% of the songs on my player have peaks > 1.  Pretty much any modern music will come out well over 1 if MP3 compressed because lossy compression is only approximate and need not have anywhere near the same amplitude as the input signal.

Interesting.  I know that 1.0 is the maximum possible value for Flac files, but I'm not sure I understand why being Mp3 encoded makes a difference, or why the peak amplitude would differ significantly.

What software do you use to do the RG analysis and tagging of the Mp3 files?

My Mp3 library is transcoded from my Flac library, to which I've run RG tags using metaflac.  I'll have to set up a test and compare the values computed on the Flac files to those computed on the transcoded Mp3 versions.

[Buschel]

Following simple example from signal theory: Take a full scale (+/- 1.0) square wave signal and cut the upper frequencies. In effect you will see overshooting/ringing (so-called Gibbs-phenomenon) that has a higher amplitude than full scale.
Just think of a lossy encoder as a tool to remove signal parts and (often) higher frequencies. When decoding/synthesizing the signal again it may have higher peak scale as the original waveform -- the overall scaling keeps the same.
As most current productions use high amplitude combined with dynamic compression (to achieve high volume/loudness), such peaks >1.0 will appear quite regularly when using lossy codecs. Generally speaking each peak>1.0 will result in clipping during playback. To avoid clipping and to keep the overall scaling constant during playback replay gain can scale all samples with 1/peak, if peak is >1.0

[Lear]

Right, I'm aware of the available settings.  I was just wondering whether the other RG values are used in the implementation or if they were ignored.

REPLAYGAIN_REFERENCE_LOUDNESS is ignored; the others are used. I hadn't heard about REPLAYGAIN_REFERENCE_LOUDNESS before, but according to this post (at Hydrogenaudio, by the author of the ReplayGain standard) it shouldn't be used.

[Chronon]

Following simple example from signal theory: Take a full scale (+/- 1.0) square wave signal and cut the upper frequencies. In effect you will see overshooting/ringing (so-called Gibbs-phenomenon) that has a higher amplitude than full scale.

What a concise and clear example.  Thank you! 

[Carson Dyle]

Do I understand correctly that Rockbox can only read ReplayGain metadata in Mp3 files from ID3v2 TXXX frames, or is it able to read the RG values that LAME writes to the LAME tag?

[Lear]

ReplayGain in TXXX and RVA2 frames are supported, but not the values in the LAME tag.