Create and evaluate probabilistic transcriptions (PT's) of speech recordings generated by mismatched crowdsourcing: by many people who don't speak the language used in the recordings.
The technique is described in this ICASSP paper, this AAAI paper, and without jargon in this Technograph article.
The theory is described in sections II.C, III.B, and V of this IEEE TASLP paper.
A stage-by-stage description is found in this Interspeech paper.
Install OpenFST, Carmel, and at least the compute-wer executable of Kaldi.
git clone https://github.com/uiuc-sst/PTgen
cd src
make
The first time you make, you'll be asked to enter the directory of OpenFST's file fst/compat.h.
This is usually /usr/local/include. If that fails, rm config.mk; make, and instead try a result from the command locate fst/compat.h.
See the subdirectory mturk.
Edit the settings file, e.g. test/ws15/settings_full.
- Ensure that the required files within that file's
$DATAexist, or can be downloaded from that file's$DATA_URL.
If needed, split the transcriptions into train/dev/eval sets.
Process the PT's: run.sh settings_full.
If run.sh can't find the executable programs of OpenFST, Carmel, or Kaldi, it prompts you for their locations,
and caches your answers in a new file config.sh, for future runs.
If you encounter errors and fix them, you can save time by starting run.sh partway through:
in your settings file, set startstage to one past the last successfully completed stage.
If you're using MCASR, instead of run.sh use run-mcasr.sh.
cd test/ws15 (or any other test directory).
../../run.sh settings-foo
The settings file in each of these tests includes
a $DATA_URL for downloading the test's data,
which is too unwieldy to store on github.
If ../../run.sh prompts you again for the locations of exes, you can just abort it with ctrl+C, retrieve those settings with cp ../../config.sh ., and retry.