This recipe splits a long utterance, containing sentences spoken in multiple languages, into multiple smaller segments.
Each split segment is expected to have speech pertaining to only a specific language.
Thus for e.g.,
if long utterance = [<music> <amharic> <english> <amharic> <english>],
then running the HMM based language ID detector on the long utterance will generate,
short utt 1 = [<music>],
short utt 2 = [<amharic>],
short utt 3 = [<english>],
short utt 4 = [<amharic>],
short utt 5 = [<english>],
The long utterances are sourced from SBS corpus. See local/make_sbs.sh
The SBS corpus is located at "/ws/rz-cl-2/hasegawa/amitdas/corpus"