= README =

 Plese note, there is now a newer, more streamlined recipe in ../s5b that we will be
 recommending going forward.

- In this recipe you can build systems for 3 different conditions:
  - IHM : independent headset microphone, run_ihm.sh
  - MDM : multiple distant microphone, run_mdm.sh
  - SDM : single distant microphone, run_sdm.sh

- Please run 'run_prepare_shared.sh' before building any system (prepares annotation and LM)

- To show the results run './RESULTS_(ihm,mdm,sdm)'

- For information about the database see : http://groups.inf.ed.ac.uk/ami/corpus/overview.shtml

= DONE =
- Refactored the scripts,
  - separate preparation of annotations + LM, which is independent on condition,
  - put 'stages' to the recipes,
  - fix the DNN recipe for LDA-MLLT features (we have no fMLLR),
  - use sclite scoring also for IHM => better logs,
  - fixed check on #wavs after downloading,
  - checked the IHM accuracy is same as in RESULTS_*!
  - defined 10h subset (now unused, 0.2-0.3 degradation at 'tri4a' if subset used for mono + tri1)

= TODO =
- update the RESULTS* files
- MDM download fails for IS1003b, IS1007d. Were those badly and should be excluded?
- wav download is slow, propose mirror for wavs? compression for wget?

= HOW TO IMPROVE RESULTS =
- Try adding pitch features,
- Try Bottlenck-DNN system like in ../../tedlium/s5/local/nnet/run_dnn_bn.sh,
- Consider use of diarization in MDM recipe, plus adding SAT training for MDM condition,