3 Familiarization

This section serves as a cursory overview of Kaldi’s directory structure. The top-level directories are egs, src, tools, misc, and windows. The directories we will be using are egs and src.

egs stands for ‘examples’ and contains example training recipes for most major speech corpora. Training recipes are available for the Wall Street Journal Corpus (wsj), TIMIT (timit), Resource Management (rm), and many others. Under each of these directories are usually a few different versions (s3, s4, s5, etc.) The highest number, usually s5, is the most current version and should be used for any new development or training. The older versions are kept for archival purposes only.

src stands for ‘source’ or ‘source code’ and contains most of the source code for programs that the training recipes call.

For each training recipe directory, there is a standard sub-directory structure. This is best exemplified in the Resource Management directory (egs/rm/s5). The top directory contains the run script (run.sh), as well as two other required scripts (cmd.sh and path.sh). The sub-directories are conf (configuration), data, exp (experiments), local, steps, and utils (utilities). The directories we will primarily be using are data and exp. The data directory will eventually house information relevant to your own data such as transcripts, dictionaries, etc. The exp directory will eventually contain the output of the training and alignment scripts, or the acoustic models.

Example directory structure