Statistical Syllabification with Supervised and Unsupervised Algorithms

1
Statistical Syllabification with Supervised and
Unsupervised Algorithms

• Shankar Ananthakrishnan
• CS562 Term Project
• December 2, 2004

2
Definitions

• A syllable is a unit of spoken language larger
  than a phoneme, smaller than a word
• Consists of three parts
• nucleus main sonorant vowel/diphthong
• onset set of consonants preceding nucleus
• coda set of consonants succeeding nucleus
• Onset and coda are optional
• Example alert ? ah l er t ? (ah) (l er t)
• Two syllables (ah) and (l er t)
• (ah) has only nucleus, no onset or coda
• (l er t) has onset, nucleus and coda

3
Usefulness

• Syllables capture higher level contextual
  information about speech than phonemes
• Can be used as acoustic units in ASR
• Sethy et. al. Syllable-based recognition of
  spoken names, ICSA PMLA Workshop, 2002
• Syllable-like units can better model
  co-articulatory dependencies
• Useful for natural speech synthesis
• Syllabic stress is an important
  linguistic-prosodic feature
• Word sense disambiguation, POS tagging

4
Syllabification

• Basic task given phoneme sequence, find the
  correct syllable bracketing
• (ah) (l er t) is correct but (ah l) (er t) is not
• How? Based on phonotactic rules of the language
• Can a machine learn these rules?
• Sure, just code them right in (NIST tsylb2)
• Downside need deep knowledge of phonology
• Or, learn them in a statistical fashion from
  large corpora (no linguistics, generalizes across
  languages)
• Labeled corpora supervised learning (ML)
• Unlabeled corpora unsupervised learning (EM)

5
Framework

• Every vowel is the nucleus of a syllable
• Noisy-channel model
• Source generates nuclei (vowels/diphthongs)
• Channel transduces nuclei to syllables
• Source output (vowel sequence) is given
• Model parameters channel probabilities p(SN)

p(N)
p(SN)
ah er
(ah) (l er t)
Source
Channel
6
Channel Model

• p(SN) p(O1,C1,O2,C2,,On,Cn N1, N2,, Nn)
• Complete channel model
• Too many parameters
• Extremely complex, intractable
• Simplification assume short term dependencies
  for onset, nuclei, and coda
• p(SN) p(O1,C1N1,N2) p(O2,C2N1,N2 ,N3)
  p(O3,C3N2,N3 ,N4) p(On,CnNn-1,Nn)
• Model is still complex
• Use graphical techniques to simplify

7
Graphical Model
Nn-1
Nn
Nn1
Cn-1
On
Cn
On1

• p(On, Cn, Nn-1, Nn, Nn1)
• Millions of parameters!
• Impossible to estimate from real data
• Graphical representation suggests good ways to
  prune model

8
Graphical Model
Nn-1
Nn
Nn1
Cn-1
On
Cn
On1

• Assume
• Nuclei generated independently of one another
• Coda independent of preceding nucleus
• Onset independent of succeeding nucleus
• Onset, coda conditionally independent given
  nucleus

9
Graphical Model
Nn-1
Nn
Nn1
Cn-1
On
Cn
On1

• p(On, Cn Nn-1, Nn, Nn1) p(On Nn-1, Nn)
  p(Cn Nn, Nn1)
• Manageable set of parameters (w/smoothing)

10
Parameter Estimation

• Supervised learning
• Training data has syllable alignments
• Maximum likelihood count-and-divide
• Unsupervised learning
• No training alignments EM!
• Similar to Eng-Jap phoneme alignment problem
• Initially assume all alignments are equiprobable
• Iteratively update channel parameters to maximize
  likelihood of observed training corpus

11
Smoothing

• Parameters are of the form p(On Nn-1, Nn)
• Sparsity parameters may not cover test corpus
• Simple interpolation
• p(On Nn-1, Nn) a1 p(On Nn-1, Nn)
  a2 p(On Nn)
  a3 p(onset Nn)
• a1 a2 a3 1.0
• Similar smoothing for coda parameters

12
Data / Tools

• CMU Pronouncing Dictionary
• Training corpus 123,777 words
• Test corpus 3,007 words
• Held-out data 2,891 words (to be used later)
• Syllabification tool NIST tsylb2
• Deterministic rule-based syllabifier
• Daniel Kahn, Syllable-based Generalizations in
  English Phonology, Ph.D thesis, University of
  Massachusetts, 1976.

13
Evaluation

• Baseline to compare against
• Alignment based on maximal onset principle
• Syllables prefer a maximum number of consonants
  in their onset and a minimum number in their coda
• t r ae n s f er ih ng ? (t r ae) (n s f er) (ih
  ng)
• Results reported in terms of
• Word accuracy how many words did the
  algorithm(s) align correctly?
• Syllable accuracy how many syllables overall did
  the algorithm(s) correctly identify?

14
Results
Table 1. Syllabification accuracy for various
algorithms
15
EM Unexpected Results
16
To-do List

• Supervised learning
• Use EM to optimize smoothing parameters on
  held-out data (possible!)
• Unsupervised learning
• Why is there a paradox in EM syllabification?
• Figure out a way to lump smoothing parameters
  along with the rest and jointly optimize w/EM
  (possible?)
• Overall goal
• Compare all techniques (including rule-based) on
  human-labeled corpus (need data!)