LOU Xiaoyan, LI Jian Research and Development Center, Toshiba China

1
LOU Xiaoyan, LI JianResearch and Development
Center, Toshiba China
Suggestions on Tone and Word Boundary of Mandarin
for SSML
2
Outline

• Tone
• Word boundary

3
Tone (cont)

• Importance
• As important as phonemes in tonal language
• Same syllables with different tones take
  different meaning
• ?(ma) ?(má) ?(ma) ?(mà)
• Sandhi phenomenon in tonal language
• ?? ni3 hao3 ? ni2 hao3
• Synthesis with correct tone help listener catch
  the meaning of speech

• Non-markup behavior
• Tone can be achieved by looking up dictionary or
  applying rules.
• Errors may occur, especially in dealing with
  sandhi

4
Suggestion on Tone (cont)

• Our suggestions
• Using Pinyin sequence as the value of phoneme
  element
• Using number 1, 2, 3, 4 and 5 standing for tone
  yin ping, yang ping, shang sheng, qu
  sheng and neutral tone in Mandarin
• Text ??(dàdou)
• Pinyin sequencetone /da 4/dou 1/
• Solution1 new tone element (optional), with
  required attribute detail
• ??
• Solution 2 new value t and ptof alphabet
  attribute in phoneme element
• ??
• 
  ??

5
Note on Tone Markup

• Possible influence on SSML1.0
• Solution 1 Tone element cannot be followed by
  other element, and can be enclosed by p, s,
  w(if defined) element
• Solution 2 phoneme element is modified, the
  relation to other elements should not change
• The tone strings given by markup cannot be
  changed
• in the text normalization step
• in the result of looking up the lexicon.
• Tone markup should be neglected, when
• Value error of tone
• Unmatched length of tone sequence

6
Outline

• Tone
• Word boundary

7
Word Boundary (cont)

• Word is the basic unit for sentence parsing and
  understanding.
• Chinese sentences are composed of sequence of
  Chinese characters without blanks or spaces to
  specify word boundaries.
• Difficulties
• Complex words, such as reduplications, derived
  words, such as ????(very easily),
  ???(immateriality)
• Proper nouns, such as location name, person name
• The ambiguous word segmentations.
• A ?? ? ? ????(Shanghai is a metropolis)
• B ??? ?? ? ?? ??(Most Shanghainese will say
  that)
• Non-markup behavior
• Determine the boundary using language-specific
  knowledge
• Errors may occur

8
Suggestions on Word Boundary (cont)

• New element w is suggested
• ??

• An optional attribute detail is also recommended
  to mark phrases
• ??????
• Here, the phrase is split into three words, and
  the number of Chinese characters of these words
  are 3, 2 and 1.

9
Suggestion on Word Boundary (cont)

• Legal values of the optional attribute detail
• Not bigger than the length of the contained text
• ??
• Default value is the length of the contained text
• ??
• When the sum of value is smaller than the length
  of the contained text, the left part is regarded
  as a word
• ??????
• The first 3 Chinese characters ???are regarded
  as one word and the left ??? are regarded as
  another word
• When the sum of value is bigger than the length
  of the contained text, this markup should be
  neglected

10
Possible Influence on SSML 1.0

• Influence on speech synthesizing steps
• Word segmentation is suggested to be done before
  parse text and analysis structure
• Relation between SSML 1.0 markups and word
  segmentation markup w (needs more discussion)
• p, s element can be followed by w element
• w element can be followed by audio, emphasis,
  phoneme, prosody, say-as, sub, voice and t(if
  defined)
• ??
• ?????

11
Thank you!