Taiwanese is a tonal language which use pitch to distinguish lexical or grammatical meaning.

The following tone sandhi rules are based upon MTL writing system. It adopts multi-syllables words that have incoorporated tone sandhi rules. Students only need to learn when the last syllable of a word need to change tone or not in a sentence.

• All the words change tone unless it is the last word of a sentence or before a punction.

  E.g. Lie karm u pid? U, goar u cidky.
  When we speak, pid, cidky and U don't change tone. The rest of words change tones based upon tone-circle pictured in this doc

• Vocabularies with backquote (e.g. aang`ee) follow back quote's rule:

  No tone changes for the word before backquote. Instead, the syllables after backquote changed to shorter and softer tones.

• Generally, nouns (include gerunds) do not change tone, unless it is used as an adjective, or measure words denoting a unit (e.g.: Taioaan-laang, cidky pid).

  For TTS, this appears to be the biggest challenge. Using POST (part of speech tagging - lexical category) would help. But this is a huge task.
  What are the existing databases out there? What models have been developed for tagging Taiwanese? Can we revise the existing POST for a jump start?
  Further, a vocabulary could be a noun or a verb. A good dataset must include these words with different meaning for training purpose.

  Meta's M2M uses different approach to solve complex tone sandhi probelms. Meta’s new AI-powered speech translation system presents a new approach for an unwritten language.

  Gemini has a detailed outlines for using transformer to tackle this issue.

  The majority of Taiwanese vocabularies are 2-syllable words. The front syllables may be derived from a syllable which changes the tone based tone sandhi rule. But as a whole, when it is used as a noun, we don't perform any further tone change. Many words contain 'ar' at the end. When creating a word with 'ar', the front syllable will change tone. E.g "niaw" -> "niau'ar" (kitty cat). But if the front syllable is a flat tone, it does not change tone when follows with 'ar'. E.g. "te" -> "te'ar" (bag).

  N	Noun
  C	Conjunction
  r	Pronoun
  

• Pronouns do change tone (goar, lie), unless specifically put a pause during speaking. (eg. emphasizing 'I': goar)

• Demonstrative adverbs/pronouns that have no tone changes are: ciaf, hiaf, zef, hef Note:
  Cit'niar si oseg`ee. where cit'niar does not change tone.

  Cit'niar oseg`ee cyn suie. where cit'niar is an ajective, hence it does change tone.

• Interrogative pronouns that have no tone changes are: tangsii, symmih, uixhøo

• Conjunctions don't change tone are : citmar, tvaf, pwntea, afnny

• Words before conjunctions that don't change tone: A'Efng gaau køq kutlat. Y si beq khix iafsi m khix. where 'gaau', and 'kix' do not change tone.

• Adjectives before 'ee' don't change tone: e.g. byle ee koniuu (vs. suie koniuu, where suie changes tone)

• Functional word - ka: When it is followed by a verb, no tone changes.

  Goar ka korng.

  Goar ka y korng. - but this one does change tone.

• Some phrases with parallel structure (verb object, verb object), there is no tone changes for the front vocabulary:

  
  e.g. phaotee khaikarng
       kofng'oe sngroe
       iusafn oansuie
       kekef hoxho
       hongthor binzeeng
       pexlau kviafiux
       lily laklag
  
  The following do change tones, since the nouns are used as adjectives:
  
  e.g: tikhaf-mixsvoax
       zekpeq-hviati
  
  

• Checkout this verb zhwlie in the following two senenteces. In the 1st sentences, it change tones, but it does not in the 2nd sentence.
  MTL ti zengphvi-bwym hongbin, zhwlie kaq cyn suykhuix.
  
  Iong MTL laai zhwlie khaq ittix.
  

• Words that never changed tone

  Note: MTL is based upon southern Taiwan's accent. Hence, some tone change like curving tone to low falling tone or short tone to long tone are avoided.(e.g. mii -> miphoe, instead of mixphoe; ciah -> ciaqpng instead of ciaxpng). Of course, people would speak with their preferred tones.

  In addition, there are colloquial and literary tones in Taiwanese. While the pronounciation might be quite differnt (e.g. zap vs. sip), tone changes rules are the same.

  For NLP, MTL has a big advantage due to it's adpotion of the concept of "voacbulary". Unlike 'harnji', the majority of the words in MTL are two syllables. E.g. 'sikoef' which stands for watermellon. And because it is a noun, there is no tone changes while speaking. Using multiple syllables for a word also overcome the ambiguity(homonyms) issues that other systems such as POJ encountered.

  hofng teq zhoef hongzhoef hongzhoef ho hofng zhoef`khix

  ---

  • Universal POS tags
  • Processing Techniques for Written Taiwanese -- Tone Sandhi and POS Tagging
  • A Knowledge Representation Method to Implement A Taiwanese Tone Group Parser
  • Processing Techniques for Written Taiwanese -- Tone Sandhi and POS Tagging
  • MTL input method installation and configuration
  • Complete guide for training your own Part-Of-Speech Tagger
  • CkipTagger