PLS for SSML

1
PLS for SSML

• Paolo BaggiaLoquendo
• Workshop II on Internationalizing SSML

2
Outline

• Brief Introduction of Pronunciation Lexicon
  Specification
• Examples of common use cases
• How to deal of Homographs in PLS
• Other issues for the workshop

3
Why Pronunciation Lexicon Specification?

• Allow to customize pronunciations (for proper
  names, locations, addresses, acronyms, etc.)
• Enrich TTS and ASR with customized pronunciations
• Complete the Speech Interaction Framework

?Read the specification athttp//www.w3.org/TR/p
ronunciation-lexicon/
4
What PLS 1.0 is Not!

• Multilingual pronunciation lexicon? the current
  specification is mono-lingual!
• Extended purpose lexicon? no syntax, no
  semantics, no morphology
• TTS-internal lexicon? too complex and rich of
  custom knowledge

PLS 1.0 is restricted to the most important and
tractable issues.
5
The PLS 1.0 Language

• PLS is an XML language lt?xml version"1.0"
  encoding"UTF-8"?gt
• The root element is ltlexicongt, with attributes
  version, xmlns, alphabet, xmllang
• It contains a collection ltlexemegts, which is
  composed of
• ltgraphemegts for orthographies/spellings
• ltphonemegts for pronunciations
• ltaliasgts for textual substitutions
• ltexamplegts for examples
• ? The order is relevant to determine the
  preferred pronunciation for TTS

6
PLS Common Usages

• Multiple pronunciations
• For ASR to accommodate speaker/regional
  variability, not native speakers
• For TTS a single preferred pronunciation will be
  selected
• Multiple orthographies (with same pronunciations)
• For ASR TTS
• Homophones (same pronunciations, different
  meanings)
• Different ltlexemegts
• Homographs (same spellings, different
  pronunciations)
• This creates troubles ? we propose a solution

7
How to Deal with Homographs

• New attribute role on ltlexemegt elementsltlexeme
  rolevalue/gt
• Values are qnames (qualified names, with a
  namespace)e.g. myvocabularyverb,
  wordnetverb
• Open to future standardization ? allows both
  proprietary values and standard ones
• More than one qname for a single ltlexemegt,e.g.
  rolewverb wpast-tense

8
Example of Homographs in PLS

• lt?xml version"1.0" encoding"UTF-8"?gtltlexicon
  version"1.0"
• xmlns"http//www.w3.org/2005/pronunciation-l
  exicon"
• alphabet"ipa" xmllang"en-GB"gt ltlexeme
  role"wdnverb"gt
• ltgraphemegtrefuselt/graphemegt
• ltphonemegtr?'fjuzlt/phonemegt
• ltlexemegt
• ltlexeme role"wdnnoun"gt
• ltgraphemegtrefuselt/graphemegt
• ltphonemegt'refjuslt/phonemegt
• ltlexemegt
• lt/lexicongt

Example sentenceI refuse to take the fridge as
a refuse.
9
SSML should to be extended!

• lt?xml version"1.0" encoding"UTF-8"?gtltspeak
  version"1.0" xmllang"en-US"
  xmlns"http//www.w3.org/2001/10/synthesis"gt
• ltlexicon uri"http//www.example.com/example.pls
  "gt
• I lttoken role"wdnverb"gtrefuselt/tokengt to take
• the fridge as a lttoken role"wdnnoun"gtrefuselt/t
  okengt.
• lt/speakgt

and SRGS too!
10
Other Issues for the Workshop

• Lexicon selection criteria in SSML
• Which phonetic alphabets?
• Current PLS 1.0 mandates the usage of IPA
  (International Pronunciation Alphabet)
• Are there other options? We need standard
  alphabets!
• Other issues?