Evaluation of GmmLatgenWrapper and its Python wrapper PyGmmLatgenWrapper

See the (intermediate) results at http://nbviewer.ipython.org/github/oplatek/pykaldi-eval/blob/master/Pykaldi-evaluation.ipynb

We evaluate the dynamic properties of decoder e.g. Real Time Factor (RTF), Latency (LAT), Word Error Rate (WER) and Sentence Error Rate (SER) based on decoding parameters.

We choose to investigate the influence of

• beam
• lattice-beam
• max-active states
• wave length

On metrics RTF, LAT, WER and SER mentioned above.

We did not experiment with the Lexicon size and Language Model (LM) complexity, which certainly influence all the metrics, but especially the LM complexity is hard to describe and visualise.

Note that decoding is language independent since the LM complexity and lexicon size is fixed. We used LM TODO with lexicon size 17000 TODO?

Technical details

The evaluation is perform using the pykaldi-latgen-faster-decoder.py and its launcher run_pykaldi-latgen-faster-decoder.sh from kaldi/egs/vystadial/online_demo directory.

TODO More describe test sets (So far I use the default from vystadial server)

We suppose the code is run from subdirectory of online_demo, so copy this repository to kaldi/egs/vystadial/online_demo.

./collect_data.sh | tee text_command.log
./parse_collect.py text_command.log pickled_tuples.txt