PREPARING ACOUSTIC MODEL TRAINING WITH SPHINX


This set of scripts are created to facilitate the creation of training and test data to use with pocketsphinx and sphinx3. So far it has only been possible semi-automate the process.

The scripts require data in PRAAT textgrids, the corresponding sound files and that the filenames have a unique identifier to identify the correspondance. 
The textgrids need to have 2 tiers: orthography and phonetic transcription. The first tier must contain the orthography and the second tier the phonetic transcription.
The scripts also assume that CMULMTK is installed. If it is installed locally, you can change the commands in train_CMU_LM.py.
 

The script is called with several arguments, e.g.

~$ python setup_sphinx_db.py path_to_top_dir path_to_grid_dir dictionary_setting


path_to_top_dir: the path to where you want the top directory to be placed. End the path with the name of the database, e.g. /home/user/Desktop/<database name>

path_to_grid_dir: the path to the directory where all the textgrids are stored.

dictionary_param: different settings can be used when you extract the dictionary from the transcriptions. The setting is used in extract_pron_dic.py where it is possible to add different extraction methods, stoplists etc.


After running setup_sphinx_db.py, you can move your sound files into their appropriate places according to the sphinx training manual in the wav/ directory. Then run

~$ python create_fileids.py abs_path_to_top_dir


to create the '.fileids files. 

Finally, you need to create the list of fillers.

The Pocketsphinx/sphinx3 database should now be configured as specified in the walkthrough on http://cmusphinx.sourceforge.net/wiki/tutorialam and ready to follow the rest of the tutorial.