1
Improving Pronunciation Dictionary Coverage of
Names by Modelling Spelling Variation - Justin
Fackrell and Wojciech Skut

• Presented by Han

2
The Problem

• The pronunciation of out-of-vocabulary (OOV)
  words is a major problem in TTS.
• Many OOV words are names.
• For English names, the orthography for names is
  highly irregular.
• Current methods of approaching this problem has
  low accuracy.
• Using hand-written or automatically learned rules
  to replace a sequence of graphemes by a sequence
  of phonemes.

3
The Challenge
4
Their Method

• Scope English surnames, forenames, street names
  and place names.
• Based on the observation that some of the words
  in the above categories have same pronunciation,
  but slightly different spelling.
• Approach learn from existing data (data-driven)
  of the rules of these variations, so that next
  time we see an OOV word, we will try to apply
  these rules and see if we can transform that word
  into an IOV word.

5
Different Orthographical Expressions for the Same
Pronunciation
6
Hypothesis

• Given a name thats not in the dictionary,
  theres about 10 chance that it DOES have a
  valid pronunciation in the dictionary. We have to
  somehow map it to a valid in-dictionary word.

7
A Hierarchical Approach
8
Two Ways of Using This Method and Their Results

• Online
• Results suggested pronunciations are good in 80
  of cases.
• Offline
• For surnames, a model trained on a 23,000-entry
  dictionary was able to add 5,000 new entries,
  increasing the coverage by about 1.

9
The Algorithm (Part I)Training

• 1) reverse dictionary (pron -gt ortho)
• 2) delete one-to-one mappings
• 3) Each pair of spellings that share a common
  pronunciation generates a set of rewrite rules,
  ri where i 0 to n, in the form of A -gt B / L _
  R

10
The Algorithm (Part I)Training
11
The Algorithm (Part I)Training

• Each rule, ri, is then evaluated on the rest of
  the dictionary to see how useful it is.
• MISS
• OOV
• DIFF
• GOOD
• And gets four scores niMISS, niOOV, niDIFF, and
  niGOOD
• From each set of rules generated by a pair, only
  one rule is chosen shortest and niDIFF 0.

12
The Algorithm (Part I)Predication

• Sort all rules by score.
• When given an OOV word, use the rule with the
  highest score that can map it into an IOV word.

13
Some Examples of Resulted Rewrite Rules
14
Some Results
15
Accuracy Test Results
16
Accuracy Test Results