FFmpeg audio dither creates unwanted DC offset on Mac

[mstgrv]

I ran a test of FFmpeg's audio dithering and I noticed that it creates a DC offset in the resulting audio file, which shouldn't be there. Doing the same dithering process with other tools produces no DC offset in the resulting audio.

Here's what I did:

1. Create a silent WAV file (44.1khz, 32-bit floating point) using iZotope RX.
2. Convert that file to 16-bit with triangular dither, using iZotope RX
3. Convert the same file to 16-bit with triangular dither, using FFmpeg. The command I used was:

   ffmpeg -i input.wav -c:a pcm_s16le -dither_method triangular output.wav
   

Because the file in step 1 is completely silent, the files created in steps 2 & 3 contain just the dither by itself.

FFmpeg's conversion results in a file with DC offset, but iZotope RX's conversion has none. The expected result is no DC offset.

I have attached all the WAV files, plus screenshots of the analysis of each file (using iZotope RX) which shows the DC offset result.

[mstgrv]

A similar bug existed in Reaper, which was fixed in 2021 ​https://forum.cockos.com/showpost.php?p=2425940&postcount=25

The developer clarified what the issue was and how it was fixed ​https://forum.cockos.com/showpost.php?p=2801990&postcount=33

The bug (fixed in v6.27) was using floor() instead of rounding when reducing floating point samples to the desired bit depth. For example if you have a variable in [-1,1] with mean=0, floor(x) will have a mean of -0.5, but floor(x+0.5) will have a mean of 0.

Perhaps this could be the fix for the issue in FFmpeg?

[Balling]

Paul that worked on this moved to his own fork of ffmpeg. Open the issue there.

​https://github.com/librempeg/librempeg

[MasterQuestionable]

͏    Did.

͏    Potential rounding error of "swresample"
͏    ​https://github.com/librempeg/librempeg/issues/11

͏    ----

͏    "-dither_method triangular"
͏    Does it happen with other methods?
͏    Or the default..?? 0?

[MasterQuestionable]

͏    Your files can be deleted.

[Balling]

Found it, and looks related to our code. ​https://gitlab.freedesktop.org/pipewire/pipewire/-/merge_requests/2040

And this literally about dither

​https://gitlab.freedesktop.org/pipewire/pipewire/-/issues/4057

[Balling]

Well, this is fixed now. There is no DC offset anymore with iZotope RX Alt+D menu. Even on triangular_hp dither.

[mstgrv]

Thank you - has the fix been released? Just trying this again with v8.1.2 and I still see DC offset

[Balling]

Replying to mstgrv:

Thank you - has the fix been released? Just trying this again with v8.1.2 and I still see DC offset

ffmpeg -i "silence 44-32.wav" -c:a pcm_s16le -dither_method triangular output.wav

produces this file that has no DC offset in iZotope RX 12.

But you can test latest master branch Windows build here: ​https://github.com/BtbN/FFmpeg-Builds/releases/download/latest/ffmpeg-master-latest-win64-gpl-shared.zip

[Balling]

Wait, I actually cannot see any DC offset on 8.1.2... Weird, maybe it is some CPU issue?

[mstgrv]

Replying to Balling:

Wait, I actually cannot see any DC offset on 8.1.2... Weird, maybe it is some CPU issue?

I'm on a Mac, if that helps track it down?

Mac M4 Pro
macOS Tahoe 26.5.2
FFmpeg 8.1.2 installed via homebrew

Just retried it again and confirming I still get the DC offset when I check in RX

[Balling]

Yes, I never checked this originally, as I thought Paul reproduced it. But it is only in ARM. Linux x86 not affected either.