SSML extensions for multilanguage usage

1
SSML extensions for multi-language usage

• Davide Bonardo

W3C Workshop on Internationalizing SSML Crete,
30-31 May 2006
2
About Loquendo

• RD of speech technology
• Over 30 years experience (from CSELT
  laboratories)
• Technologies
• TTS (text to speech)
• ASR (automatic speech recognition) SV (Speaker
  Verification)
• Solutions
• Easy integration of speech technologies
• Speech servers (MRCPv1 v2 protocols)
• Speech platforms (VoiceXML CCXML interpreters)
• Embedded solutions (for many OS and devices)

3
Ideas for SSML extensions

• ltsay-asgt element
• Extension of the values for the interpret-as
  attribute
• New element
• lttokengt

4
Proposal 1 ltsay-asgt extension (1/3)

• Problem
• How to interpret a part of an input text
• Different contexts of dialog require different
  interpretations
• The interpretation could be language dependent
• Many contexts could be defined sms, e-mails,
  news, application for rescue operations,
• The TTS engines may use context information to
  activate the best configuration for
• reading acronyms
• abbreviation expansions
• using customized prosodic phrasing
• activating a special reading style

5
Proposal 1 ltsay-asgt extension (2/3)

• Proposal
• To extend the interpret-as attribute with new
  values, for instance
• sms
• e-mail
• news
• banking
• navigation

6
Proposal 1 ltsay-asgt extension (3/3)

• Examples
• lt?xml version"1.0" encoding"ISO-8859-1"?gt
• ltspeak version"1.0" xmllang"en-US"gt
• I call you asap.
• ltsay-as interpret-as"sms"gt
• I call you asap
• lt/say-asgt
• lt/speakgt
• lt?xml version"1.0" encoding"ISO-8859-1"?gt
• ltspeak version"1.0" xmllang"en-GB"gt
• ltsay-as interpret-as"sms"gt Mtfbwu lt/say-asgt
• lt/speakgt

7
Proposal 2 New element lttokengt (1/3)

• Problem 1 the activation of the correct language
  knowledge at the specific point of the text
• xmllang attribute is currently available in
  ltspeakgt, ltvoicegt, ltpgt and ltsgt elements
• The behavior for the engine could be different
• In the root ltspeakgt element, xmllang defines
  the language of the whole document, but for the
  engine it involves the selection of a voice
• In the ltvoicegt element, it is an important
  recommendation in order to load the correct voice
• In the ltpgt and ltsgt elements, it is mainly a
  language information and the engine, if able to
  do this, can use the same voice but a different
  language knowledge (e.g. phonetic mapping)
• Problem 2 it could be necessary to specify a
  language change for a text unit smaller than a
  sentence.

8
Proposal 2 New element lttokengt (2/3)

• Proposal
• To introduce a new element lttokengt
• To extend the use of xmllang attribute to the
  lttokengt element
• Advantages
• It is a generic element
• It is extensible
• Without attributes, it could be used to give
  information on the segmentation, where needed.
• With other attributes, it could specify new
  information for the token (i.e. part of speech)

9
Proposal 2 New element lttokengt (3/3)

• Examples
• lt?xml version"1.0" encoding"ISO-8859-1"?gt
• ltspeak version"1.0" xmllang"en-US"gt
• The movie is the product of Italian comic
  sensation Roberto Benigni, who wore three hats
  for "La vita è bella" director, co-writer, and
  star.
• lt/speakgt
• lt?xml version"1.0" encoding"ISO-8859-1"?gt
• ltspeak version"1.0" xmllang"en-US"gt
• The movie is the product of Italian comic
  sensation lttoken xmllang"it-IT"gtRoberto
  Benignilt/tokengt, who wore three hats for lttoken
  xmllang"it-IT"gt "La vita è bella"lt/tokengt
  director, co-writer, and star.
• lt/speakgt

10
Conclusions

• Proposal 1
• To increase the number of interpret-as values
  with the identification of new context of speech
• Proposal 2
• To introduce a new element to define some
  specific information (i.e. the language) for a
  single word, or phrase and so on.