Keep in mind, that Janus is still under development, and that we can't guarantee that every script that is listed here will work correctly forever. Sometimes, a program changes faster than its supporting documentation.
The scripts were all made for the tutorial, they are not meant to work for any task in any environment. The script can be run unmodified in the do-it-yourself environment of this tutorial. None of the scripts uses any external script or library function. They are all complete. There is a big Janus script library which contains many scripts for many jobs and environments. These are not part of the tutorial. Once you know how to work with Janus you will find it much easier to use the big Janus library or to develop your own scripts.
Okay, here come the scripts:
Given a list of utterances together with their transcriptions, create a simple Janus database object.
Initialize a context-independent system, i.e. create an initial environment, create architecture description files.
Fire up a Janus process for a newly created environment, using acoustic parameters from a generic recognizer.
Use the generic acoustic models recognizer to run an alignment on the training database and write label files.
Compute a first LDA transformation matrix (and some side dishes, like class-counts) for a context independent system.
Run a forced-aligment training iteration on all training data along the previously saved labels, and store the optimized weight files.
Extract the recognition vocabulary from the dictionary.
Compute a simple unigram/bigram language model based on the training data.
Run the decoder on the test set using the previously trained weights. Here we are only run a very simple one-pass decoding.
Compute new codebooks with k-means after extracting sample vectors. Write new weights and description files.
Train a couple of iterations not doing forced alignment, but using the previously written label files.
Create a first unclustered context-dependent environment. Compute polyphone lists and write the needed description files.
Train the context-dependent system, this is similar to the "trainAlongLabels" script, only the startup differs.
Create a list of phoneme classes that can be used for the decision-tree clustering of polyphone contexts.
Cluster the polyphone contexts into fewer contexts, and create a separate codebook for each new cluster. Write the corresponding description files.
Compute an LDA transformation matrix a second time, this time using the context-dependent classes.
Compute new context-dependent codebooks with k-means after extracting sample vectors. Write new weights and description files.
Run a more sophisticated test on the test data, using multiple pass search and lattice rescoring.