In todays session we will take the context independent recognizer resulting from last session and build our first context dependent system. For this purpose you will create a context dependent environment, then train a few iterations for the new architecture, find linguistic motivated questions about the context of a phoneme. After that you will compute another LDA, initialize the codebooks of your context dependent system by extracting samples, running kmeans, and train along labels. Finally the resulting reognizer is evaluated with a multi-pass test.

Context-dependent environment

Task 17: Change your dictionary by adding "Word Boundary" tags. You can use the following command: cat dict | perl -pe 's/} {([^ ]*)/} {{$1 WB}/g' | perl -pe 's/([^ }]+)}\n/ {$1 WB}} \n/g' > convertedDict . Create a directory step10. Build a context-dependent environment by following Janus Tutorial Step10 . Don't forget to adapt all following scripts in order to read your new convertedDict from step10 - not from step1. Then create a directory step11 and train a few iterations for the new architecture by following Janus Tutorial Step11 .

Question file, clustering

Task 18: Define a question set and running the clustering algorithm by follow Janus Tutorial Step12 . Adapt the question set to our phoneme set! Discuss possible phonetic questions with your lab member - each member may try a different solution and you can compare the results!

LDA, samples, K-means, training

Task 19: Compute a new LDA for the context dependent models by following Janus Tutorial Step13 . After that initialize the codebooks by extracting samples and running K-means - see Janus Tutorial Step14 .

EM-Training along Labels and multi-pass test of the resulting recognizer

Task 20: Train the recognizer along the existing labels and evaluate the latest recognizer using a multi-pass test by following Janus Tutorial Step15 .

Last modified: Wed Apr 18 15:42:53 EDT 2001
Maintainer: tanja@cs.cmu.edu.