ParCzech4SpeechModeling

Recognition of ParCzech4 using WhisperX

Use the following scripts:

• recognize.py - For given links downloads .tar files, extracts them into the common directory and runs the recognition for a recognizers specified in recognize.yaml. For debug purposes you can set n_debug to a number greater than 0, this corresponds to the number of files in the first link that will be recognized by all recognizers in the config. Required parameters are links and output_folder.
• recognize_parallel.py - This is a wrapper for recognize.py that given all links creates n_jobs slurm jobs with recognize.py script with its own config recognize_parallel.yaml. Recognition is done on the GPU cluster. Use n_debug for debugging purposes, it will be propagated to recognize.py. The default number of slurm jobs is specified by n_jobs in the config, with -1 corresponding to the number of links (each recognize.py job will process one link). Required parameter is job_name that will be used as a prefix for the slurm jobs.

The results will be stored in the output_folder specified in the recognize_parallel.yaml, that will be propagated to recognize.py. As for now it is /lnet/troja/work/people/stankov/parczech4speechmodeling. There you can find:

• .tar files for each downloaded quarter.
• audioPSP-meta.audioFile.tsv and audioPSP-meta.quarterArchive.tsv, two files with the metadata about the original ParCzech4 dataset (audio information only).
• audio directory with all audio files and transcripts. Relative path to the audio can be find in audioPSP-meta.audioFile.tsv in the column filePath. The recognized texts will share the same relative path and will be stored in .tsv files. The suffix in the naming displays the recognizer type.