A handful of scripts I put together while evaluating online decoders. They are not particularly clean but bear with me.
The online decoders use nnet2, so any models trained with nnet1 will need to be converted. The script for this is utils/nnet1-to-nnet2.sh. More documentation is available within the file. If used correctly, running this script will result in a .mdl file suitable for use in our online decoding setups.
The main script is recognize.sh, which decodes audio and evaluates the results for a given decoder setup. The point of this script is to use host dispatch to run multiple different jobs in parallel, where each job runs with different values for the arguments passed to the online recognizer. It is currently written to accept multiple different values for the --max_active, --lattice_beam, and --beam parameters, but hopefully the idea should extend to other parameters as well. This script takes a configuration file as an argument and outputs the results of decoding into a file called results.txt. For example, to run the arabic decoder you would use:
./recognize.sh conf/gale.conf
If you have no interest in running multiple jobs and instead want to run a single job with a single set of parameter values, it would probably be easier to look at decode_and_evaluate.sh, which contains the commands necessary for decoding and evaluating for a single job. It is currently written as a subroutine of the larger recognize.sh script but it should be simple to adapt it to run as a standalone script.
In addition to the command-line decoding utilities provided above, there is also a web-based decoder. The web-based decoder integrates with the Gstreamer plugin for Kaldi.
For detailed information on getting that set up, see
https://github.com/alumae/kaldi-gstreamer-server
and
https://github.com/alumae/gst-kaldi-nnet2-online
This setup is already downloaded and compiled at /data/sls/scratch/nbravo/yifans_version
For an example implementation of a UI, see https://github.com/yifan/dictate.js