4.5 Machine learning of phonological rules

1
4.5 Machine learning of phonological rules
Presenter ??? (Ester)

• Machine learning systems automatically induce a
  model for some domain, given some data and
  potentially other information.
• Supervised algorithms are given correct answers
  for some of the data and use these answers to
  induce generalizations to apply to further data.
• Unsupervised algorithms work only from data, plus
  potentially some learning biases.

2
Learning rules (cont.)

• Ex Johnson (1984)/Touretzky et. al (1990) learn
  SPE (Sound Pattern of English (SPE)-style rules
  from a corpus of input/output pairs.
• Ex Gildea Jurafsky (1996) specialize a
  learning algorithm (based on OSTIA algorithm) for
  a subtype of FSTs (called subsequential
  transducers) to learn two-level phonological
  transducers from a corpus of input/output pairs.
• Required learning biases Faithfulness and
  Community
• -- Faithfulness underlying segments tend to
  be realized similarly on the surface.
• -- Community similar segments behave
  similarly.
• If SPE-style rules can be implemented as FSTs
  automatically, why learn the FSTs directly?

3
4.6 Mapping text to phones for TTS

• Text-To-Speech (TTS)
• Map orthography to phonetic transcription
• Add in prosody
• Map phonetic transcription prosody to acoustic
  signal
• Pronunciation dictionaries are used for both
  text-to-speech (TTS) and automatic speech
  recognition (ASR) systems. They give the
  pronunciation of words as strings of phones,
  sometimes including syllabification and stress.

4
Pronunciation dictionaries (cont.)

• A list of words and their pronunciations
• No morphological or phonological rules
• Three large pronunciation dictionaries
• Designed for ASR, but can be adapted for speech
  synthesis.

5
ARPAbet phone set and corresponding IPA symbols
6
Problems for TTS and ASR dictionaries

• Homographs (two distinct words spelled the same)
• TTS dictionaries essential
• ASR dictionaries usually ignored
• (2) Proper names dictionaries often dont
  include many proper names, e.g. Dr. , 2/3.
• (3) Function words highly variable
  pronunciations, e.g. and, I, a, the, of, etc.
• One significant difference between TTS and ASR
• --TTS dictionaries dont represent dialectal
  variation
• --ASR dictionaries represent dialectal
  variation

7
Beyond dictionary lookup Text analysis

• Text analysis
• The text-analysis component of a
  text-to-speech system maps from orthography to
  strings of phones. This is usually done with a
  large dictionary augmented with a system (such as
  a transducer) for handling productive morphology,
  pronunciation changes, names, numbers, and
  acronyms.

8
Problems for pronunciation modeling

• Names
• -- Impossible to list all proper names in
  advance.
• -- Come from any language and have variable
  spellings.
• -- Include not only peoples names but also
  company names and product names.
• -- 21 of 33 million words of AP newswire
  were names.
• Morphological productivity e.g. names, acronyms
  and new words are often inflected.
• Numbers with different possible pronunciations
• -- Serial, combined, paired, hundreds,
  trailing unit

9
FST-based approach

• Five components
• 1. large morpheme pronunciation dictionary,
  encoded as an FST
• 2. FSAs for morphology (possible sequencing of
  morphemes)
• 3. FSTs for morphophonology (like spelling change
  rules), e.g., pronunciation of -s in different
  contexts
• 4. heuristics and letter-to-sound (LTS) rules/
  transducers for names and acronyms
• 5. default LTS rules/transducers for other
  unknown words

10
Architecture

• Lexical (underlying), intermediate and surface
  levels all contain two tapes, one for orthography
  and one for pronunciation, separated by , e.g.
  ck.
• Lexicon-FST composed of two-level lexicon
  (lexical level and intermediate level) plus
  FSAs/FSTs for morphology (two levels separated by
  the )
• e.g. PL? sz (see Fig. 4.21-4.23)
• The automaton adds the morphological features
  N and PL at the lexical level, and also
  adds the plural suffix sz at the intermediate
  level.
• FST1... FSTn orthographic rules (or spelling
  rules) and phonological rules, all run in
  parallel so as to map between this intermediate
  level and the surface level.

11
Architecture (cont.)

• Text-to-speech applications Mapping between
  lexicon and surface form for orthography and
  phonology simultaneously, e.g. text generation,
  reading text out loud applications

12
Names

• Liberman Church (1992) attempt to handle most
  frequent 250,000 name tokens
• -- Dictionary of pronunciations of 50,000
  names 59
• -- Stress-neutral suffixes (-s, -son,
  -ville) 84
• -- Name-name compounds and rhyming
  heuristics 89
• -- Prefixes, stress-changing suffixes and
  suffix- exchanges ??
• -- Letter-to-sound (LTS) rules for the
  remainder.

13
4.7 Prosody in TTS

• Prosody operates on longer linguistic units than
  phones, and is the study of suprasegmental
  phenomena.
• Phonological aspects of prosody prominence,
  structure and tune.
• Prominence stress (lexical and sentential) and
  accent
• -- lexical stress e.g. table (stress on the
  first syllable)
• -- accent
• ? unaccented e.g. function words (there,
  the, a)
• ? larger accent patterns e.g. new truck
  (accent on the right word)

14
Phonological aspects of prosody

• (2) Prosodic structure some words group
  naturally together and some words have a
  noticeable break or disjuncture between them.
• Prosodic phrasing
• -- Intonational phrases larger prosodic units
• e.g. I wanted to go to London, but could only
  get tickets for France.
• -- Intermediate phrases lesser prosodic
  phrase boundaries
• e.g. I wanted to go to London

15
Phonological aspects of prosody (cont.)

• (3) Tune intonational melody of an utterance
• Intonational tunes include pitch accent
• Pitch accents occur on stressed syllables and
  form a characteristic pattern in the F0 contour.

16
English pitch accents (Pierrehumbert 1980)

• H high (on a stressed syllable)
• L low (on a stressed syllable)
• LH rise, starting on a stressed syllable
• LH rise, ending on a stressed syllable
• HL fall, ending on a stressed syllable
• (HL apparently not needed)
• e.g. oh, really ? excited version (LH)
• ? sceptical version
  (LH)
• ? angry version (L)

17
Other components of the English system

• ToBI model has two phrase accents (L- and H-) and
  two boundary tones (L and H)
• Phrase accents occur at an intermediate phrase
  boundary.
• ? L-
• ? H-
• Boundary tones are used at the ends of phrases
  to control whether the intonational tune rises or
  falls.
• ? L
• ? H

18
Phonetic or acoustic aspects of prosody

• Prominence Prominent syllables are generally
  louder and longer than that non-prominent
  syllables.
• Prosodic structure Prosodic phrase boundaries
  are often accompanied by pauses, by lengthening
  of the syllable just before the boundary, and
  sometimes lowering of pitch at the boundary.
• Tune Tune is manifested in the fundamental
  frequency (F0) contour.

19
Prosody in speech synthesis

• Major task for a TTS system
• --To generate linguistic representations of
  prosody
• --To generate acoustic patterns which will be
  manifested in the output speech waveform
• Output of a TTS system
• A sequence of phones, each of which has a
  duration and an F0 (pitch) value.
• --Duration of each phone dependent on
  phonetic context
• --F0 value influenced by lexical stress,
  accented element, intonational tune (e.g. a
  final rise for questions)

20
The FESTIVAL speech synthesis system

• sample Do you really want to see all of it?

Fig. 4.26 The F0 contour of the FESTIVAL
21
4.8 Human processing of phonology and morphology

• The distinction of representation via rules from
  representation via lexical listing. ? wrong
• Data-driven models of morphological learning and
  representation e.g. English past tense -ed
• --Connectionist model regular morpheme -ed
  emerges from its frequent interaction with other
  forms.
• --Dual processing model regular forms like
  -ed are represent as symbolic rules, but
  subregular examples (broke, brought) are
  represented by connectionist-style pattern
  associators.