Speech:Summer 2012 Future Tasks


 * Home
 * Information

Potential Future Tasks
There is a still a lot to be done with this project. I have identified several possible tasks that future groups could work on to improve the Speech project:
 * Modify the parseDictionary.pl and dictionary.pl scripts. These scripts are not very efficient.  It takes quite a long time to build a custom dictionary.  Right now parseDictionary.pl builds a unique list of words and passes each word to dictionary.pl.  dictionary.pl then runs through the entire master dictionary for each word.  Limiting the number of passes through the master dictionary would greatly speed up this process.
 * Improve the genTrans.pl script. Right now the master transcript has placeholders for noises that aren't words like [Vocalized-Noise], [Background-Noise], etc.  These tags are being converted into actual words for recognition.  The script could be improved to ignore these tags so Sphinx doesn't try to process them as spoken words.
 * Listen the audio files for words that were added to the dictionary. I added about 200 words that were not in the dictionary.  Either these were places, slang, or mispronounced words.  I used my best guess to determine the correct phones to use to pronounce these words.  However, I could be wrong and these words should be verified by listening to the actual audio file.
 * Make a script that improves the process for adding missing words to the dictionary. Right now it is somewhat of a manual process.  Currently a list of words needs to be built and added to the end of the dictionary.  Using cat  | uniq >>  will put the correct words in place.  However a script could be used to tweak the order used.
 * Look at streamlining the scripts used into possibly one script that creates the experiment, parses the transcript, makes the required sph files and does all the other tasks needed to build the acoustic model. Many of these scripts are modified variants of genTrans.pl.  A lot of these functions could be combined into one script.