QUOTE (C.R.Helmrich @ Feb 22 2013, 16:38) <{POST_SNAPBACK}>
I also heard similar issues on other items starting with a fade-in (the first half-second or so was apparently coded with SILK). It was either at 24 or 32 kbps, official v1.1 binary. Makes me wonder: which stereo tools are available for SILK? Can it do "true" stereo with a downmix+residual (or M/S) or just some kind of time-domain intensity stereo?
SILK uses downmix+residual, aka MS stereo in the signal domain (unlike CELT and Vorbis that do MS after band normalization). That's what explains some of the stereo artefacts it sometimes causes. That being said, it usually sounds good on "normal" stereo speech that doesn't have too much channel separation. I recently checked in some changes to the detector in the exp_analysis branch that improves the decision code and adds the possibility of using look-ahead (up to 2 seconds) on that decision to make it better.