[Discussion] Input samples granularity

First I'd like to thank to all the team for this amazing project. :+1:

I'd like to implement a [GStreamer](https://gstreamer.freedesktop.org/) element wrapping `spleeter`. To do so I'd need to know exactly how much input `spleeter` needs to be able to produce a good result.

Should I buffer the whole song and send it to `spleeter`as whole? Or can I split it in smaller fragment (which size?) and aggregate `spleeter`'s outputs to build the result?
The latter would reduce buffering and would make it easier to use `spleeter` in real time pipelines.