CSPP 56553 - Artificial Intelligence
Homework #6: Due March 3, 2004
Through this assignment you will:
- Explore Weighted Automata/Markov Models as a mechanism for reasoning with uncertainty over time.
- Experiment with the use of these models for pronunciation modeling and
In lecture we discussed a model of the pronunciation of the word about
that had been extracted from the Switchboard corpus, a collection of
conversational telephone speech. Here we consider pronunciation from
a different data source - TIMIT - a phonetically structured corpus of
read speech. For TIMIT, participants were asked to a read back a set
of sentence prompts. These sentences were constructed to cause each
phoneme to appear in as many contexts as possible.
Below, you will see a set of pronunciations for the word "permanent"
automatically extracted from close manual phonetic transcriptions of
the TIMIT recordings. We will construct a weighted automaton model
of this word and use it to perform some calculations.
- pcl p er m ix nx eh n
- pcl p er m ix n ih n tcl t
- pcl p er m n ah n tcl t
- pcl p er m ix n eh q
- pcl p er m ix nx ix q
- pcl p er m n ih n tcl t
- pcl p er m ax nx ix n tcl t
- pcl p er m ah n eh n q
Identify the states and the legal transitions between states.
Compute the weights (transition probabilities) for each transition
in your automaton, based on the small corpus of pronunciations.
What is the probability of the pronunciation "pcl p er m ax nx ix n q"
according to the model? ('q' represents a glottal stop; it's not a typo)
Based on this automaton, what is the most probable pronunciation?
What is its probability?
Implement the Viterbi algorithm.
Apply your implementation to either the "tomato" or "about"
automata. Demonstrate the option of the algorithm on
Note: You only need to return the maximum
probability; you do not need to return the path (unless you
want the extra challenge).