[egs] Minor fixes to x-vector based diarization/speaker verification recipes

[entn-at]

• Add missing data prep scripts for MUSAN for callhome_diarization
• Copy vad.scp (required by vad_to_segments.sh for SRE data) and segments (required by extract_xvectors.sh) to *_cmn data folders after prepare_feats.sh. The missing segments file also makes the call to fix_data_dir.sh after prepare_feats.sh fail.
• Fix check before create_split_dir.pl

[david-ryan-snyder]

Thanks! Is this ready to merge, @entn-at?

[entn-at]

Yes, it's ready to merge! I ran the entire recipe (DER: 7.47%, but I didn't use SWB2-Ph1)

[david-ryan-snyder]

Thanks! @danpovey, could you merge this when you get a chance?

@entn-at, is the DER: 7.47% using the tuned stopping threshold or the oracle number of speakers?

[entn-at]

The DER was 7.47% using the oracle number of speakers. Using the tuned stopping threshold it was 8.66%. Both are higher than the ones given in run.pl, but I assume that's because I didn't use Switchboard-2 Phase-1 as part of the training and therefore had fewer speakers (4623) and utterances for training the x-vector model.

[david-ryan-snyder]

Cool, thanks for the info.