[Speex-dev] Feature request: rewindable resampler

[Alexander E. Patrakov]

Hello.

As you probably know, PulseAudio uses the resampler from libspeexdsp by 
default. As a PulseAudio contributor, I have a feature request.

As you can see from old publications [1,2] by Lennart Poettering, 
PulseAudio has a "timer-based scheduling" feature which is now active by 
default. PulseAudio attempts to use as high latency as possible 
(sometimes up to 2 seconds) in order to save power and reduce the chance 
of dropouts. However, in order to stay responsive to events such as 
software-volume changes or a new stream appearing, PulseAudio relies on 
the ability to rewind over already-buffered but not-yet-played sound 
data and replace them with a new version.

Also I must say that PulseAudio client API includes rewinds - i.e. any 
client can say "please forget N samples that I wrote before, here is an 
updated version".

Unfortunately, we cannot pass such rewind requests through a resampler, 
simply because no open-source resampler library has an API to do this. 
I.e. there is no way to forget N last input samples, replace them, and 
figure out which output samples should be replaced with what. Currently, 
in this situation, we just reset the resampler, which is wrong.

Also there are situations where it ends up wanting to reask for N last 
output samples corresponding to a modified input where the input 
modification can be produced on-demand and logically extends to negative 
infinity in time. Usually, this happens when software-volume is applied 
before the resampler and a volume-change event occurs. The same kind of 
bug then happens: we cannot do anything better than resetting the 
resampler. End result: users complain, see e.g. [3].

Could you please consider designing and implementing new speexdsp API 
functions that would allow PulseAudio to get rid of these bugs?

In fact, because this request is, in my opinion, specific to PulseAudio 
and ALSA (in ALSA - because of the need to implement snd_pcm_rewind() 
for the rate plugin), I considered writing a new resampler from scratch 
with this feature. However, I will almost certainly not start within the 
next 3 months due to my dayjob and because of other TODO items, and 
David Henningsson has hinted me to talk to you as an upstream of speex, 
because, in his opinion, this "rewind" feature might be useful for other 
applications. Hence this email :)

[1] http://0pointer.de/blog/projects/pulse-glitch-free.html
[2] 
http://lac.linuxaudio.org/2010/download/lennarts-talk-auf-der-lac-2010.pdf
[3] https://bugs.freedesktop.org/show_bug.cgi?id=50113

-- 
Alexander E. Patrakov