Next: Suggestions for labs Up: Lecture plans Previous: Topic: Introduction to

Topic: Learning to read with NETtalk

The problem

Convert English text into the vowel and consonant sounds (phonemes) that make up speech. For example consider the following words: misled, hors d'oeuvres, and bedraggled.

Traditional Approach

  1. Create if-then rules to encode all the regularities in the language.
  2. Maintain a database of exceptions to these rules.
  3. Build a production system to resolve conflicts.

For example, a ``c'' can either be pronounced like an ``s'' as in center, icy, and city or a ``k'' as in cat and crumb. Here is a rule that might encode this difference:

if a ``c'' appears before an ``i'', ``e'', or ``y''
then pronounce it like an ``s''
else pronounce it like a ``k''
Exception: celtic

Network Approach

Allow a system to learn how to pronounce words by giving it lots of examples and correcting its mistakes.

  1. Choose an architecture.
  2. Choose a representation for the input and output.
  3. Determine a training set. 16,000 words were chosen randomly from the 20,000 words in Webster's Dictionary.
  4. Train the system. Give the network each word in the training set several times, and indicate the errors for learning.
  5. Test the system. The network was able to respond correctly to 90%of the 4,000 remaining words from the dictionary.

Advantages of the Network Approach

Transcript from the NETtalk audio tape



Next: Suggestions for labs Up: Lecture plans Previous: Topic: Introduction to


meeden@cs.swarthmore.edu