[pulseaudio-discuss] [PATCH 0/4] Add support for libsoxr resampler

[Alexander E. Patrakov]

14.11.2014 13:37, Andrey Semashev wrote:

> As a side note, I think speex-float-1 is a rather poor default as well because
> of its quality. Anyone who cares about the sound he gets will likely change it
> to something like speex-float-3 or 5 anyway. IMHO.

Initially I also thought so and argued for changing the default. But 
then I applied a psychoacoustical model and found that on real music 
that you can buy or download, due to masking, the difference between 
speex-float-1 and ideal resampler is below the threshold where a human 
can notice. On artificial testcases, yes, speex-float-1 does produce 
audible distortions, but who cares? Nobody listens to sine waves.

The prior research is available at the following locations:

http://lists.freedesktop.org/archives/pulseaudio-discuss/2014-August/021362.html 
  (initial attempt to measure distortions and judge their audibility by 
means of a psychoacoustical model, on sine waves)

http://lists.freedesktop.org/archives/pulseaudio-discuss/2014-September/021811.html 
  (attempt to take non-ideal listening conditions into account, but 
still on sine waves)

http://lists.freedesktop.org/archives/pulseaudio-discuss/2014-October/021911.html 
  (the latest results, on real music files - going away from pure tones 
dramatically changes the picture)

In order to defend your "speex-float-1 is too bad" statement, you need 
to refute at least one of my bases:

1. The psychoacoustical model described in 
http://www.mp3-tech.org/programmer/docs/6_Heusdens.pdf

2. The method of collecting a resampler response to sine waves.

3. The implicit assumption that, when a signal with non-trivial spectrum 
is resampled, the power of distortions at each target frequency adds up 
linearly. In other words, if two source frequencies contribute to 
unwanted distortions at one target frequency, these contributions are 
uncorrelated.

-- 
Alexander E. Patrakov