Extending Gentle Aligner

Week 1 Progress

Generating HCLG.fst and Acoustic model from Voxforge_ru (Russian ASR model) recipe

• Setting up Case HPC account, tried installing Kaldi on case cluster. Switched to AWS computing cluster ec2 to enable quick development.
• Installed Kaldi on AWS cluster, downloaded voxforge_ru recipe on the cluster and executed the recipe.

• Week 1 Task: Running voxforge_ru recipe, generating ASR model. Figuring out how to extract timing information from the generated model.

• In order to install Kaldi, all dependencies must be satisfied. I faced an issue with MKL installation. It existed but it’s path was not known to /.configure in ‘src’ directory of Kaldi files.

  • Use ‘whereis MKL’ to find the right path of the library Or look into bin/lib Or home/../intel/compilers_and../../mkl

• Another library ‘srilm’ required for running the russian recipe successfullly needed to be downloaded manually.

• On running run.sh, feats.scp was not created and model was also not generated.

  ERROR: FstHeader::Read: Bad FST header: standard input

  In case, you are facing the same error, copy arpa2fst from src/lmbin to /usr/bin manually. error thread

• These steps generated the model and the decoded graph in exp/mono or exp/tri

• Next task is to figure out how to extract timing information from the graph stored in exp/tri i.e. final.mdl week 2
• main page

Tools: Kaldi, Python, C, Bash Scripting

Link to GSoC Project Repository