Proper Noun Pronunciations

This directory contains databases and software to generate multiple pronunciations of proper nouns from their text-only spellings, a project in collaboration with Texas Instruments, Inc.

Strategies:
- Neural networks
  - Boltzmann machine
  - Multilayer perceptron
- Decision trees

Deliverables:
- Surname pronunciation databases developed for training and evaluating the system.
- Demo and software for the neural network and decision tree systems.

Publications:
- Annual technical report submitted to our sponsors in 1997
- An Advanced System to Generate Pronunciations of Proper Nouns --- a paper and poster presented at ICASSP '97
- The ICASSP '97 poster presentation in HTML
- An IEEE SAP paper submitted for review
- Automated Generation of N-Best Pronunciations of Proper Nouns --- a paper and poster presented at ICASSP '96
- A technical report submitted to our sponsors in 1996

Evaluations:

The evaluations reported below pertain to the Boltzmann machine neural networks and the statistical decision tree software being developed at ISIP.

For training, both systems are fed the training dictionary of name-pronunciation pairs. Each system converts the data to a character string corresponding to each phoneme in the pronunciation --- an n-tuple of characters. For instance, the dictionary entry of "Matt m @ t _" is converted to

       _ _ M a t      m
       _ M a t t      @
       M a t t _      t
       a t t _ _      _

for a context size of 5. During training the system learns the correspondence between such letter n-tuples and the corresponding phoneme.

During testing, the system converts the input names to similar n-tuples of the context size and probabilistically generates the likely choices for the corresponding phoneme. The phoneme error rate refers to the percent error in the generation of accurate phonemes.

The pairs of the letter n-tuples and generated phonemes are now formatted to revert to the original name-pronunciation format. The name error rate refers to the percent error where the generated pronunciations for the surname do not match with the reference pronunciations.

Since many surnames have multiple reference pronunciations, and the system also generates multiple pronunciations for each surname, the performance measurement involves a many-many map between the reference and hypothesis pronunciations. Therefore we have devised three categories to properly evaluate the performance of the system:

all correct:	all the reference pronunciations for that name are covered by the hypothesis pronunciations generated by the system.
some correct:	only some of the reference pronunciations for that name are covered by the hypothesis pronunciations generated by the system.
no correct:	none of the reference pronunciations for that name are covered by the hypothesis pronunciations generated by the system.

The no correct category indicates the name error rate of the system. A few evaluation results are summarized below:

128 names, each four letters long, that we often use for preliminary experiments with neural networks as well as decision trees

System	Phoneme error rate	Name error rate
Boltzmann machine	11.13%	35.94%
Decision tree	4.10%	14.06%

4-letter names --- used for preliminary experiments with neural networks as well as decision trees
- Training set
- Test set

System	Phoneme error rate	Name error rate
Boltzmann machine	20.76%	52.13%
Decision tree	17.52%	44.85%

The full data set, with 18494 surnames and 24040 pronunciations. The training set consists of 15000 names, and the test set is the remaining 3494 names.

System	Phoneme error rate	Name error rate
Boltzmann machine	37.88%	70.44%
Decision tree	-na-	46.15%