##Notes about Kaldi It is recommended that you review the [Kaldi documentation] (http://kaldi-asr.org/doc/) before you begin, especially if you intend to modify the compiled model included on Sourceforge.
The models linked to from this repo are trained on public media content from NPR and many podcasts. This software is intended specifically for the transcription of public media content, though anyone is welcome to use it for content of their choosing.
Create a directory to store results from sclite, e.g. from your main dir, do mkdir results
Usage
Run kaldi speech recognition on directory of wav files: python run_kaldi.py [EXPERIMENT-DIR] [WAV-DIR] Run evaluation: python run_sclite.py [EXPERIMENT-DIR] [RESULTS-DIR] [REF-FILE-PATH]
To view the summary results, open [RESULTS-DIR]/[EXP-RESULTS-DIR]/[EXP-NAME]_results.sys
##Building on the model You can use the current model as is, or add your own lexicon or language model.
You can add new words to the lexicon by editing exp/dict/lexicon.txt and running sh prep_lang_local.sh exp/dict exp/tmp_lang exp/lang.
Use sh add_grammar.sh [LM-FILEPATH] to create a new language model based on an LM textfile and to update the overall model.
Acknowledgements
The development of this software was funded by an Institute of Museum and Library Services Research Grant to WGBH for the “Improving Access to Time Based Media through Crowdsourcing and Machine Learning” project.