UNIT 4

Phonetics
Introduction
1. Debate Context

• The “whole language” vs. “phonics” debate in teaching reading seems modern but
mirrors an older historical debate in writing systems.

2. Historical Writing Systems

• Earliest independently invented writing systems (Sumerian, Chinese, Mayan) were

mainly logographic: one symbol = a whole word.

• Even early logographic systems included syllabic or phonemic elements where

symbols represented sounds.

• Example: The Sumerian symbol for “ration” (pronounced ba) could also function
purely as the sound /ba/.

• Modern Chinese, though mostly logographic, uses sound-based characters for foreign
words.

3. Development of Purely Sound-Based Writing Systems

• Sound-based writing can be:

o Syllabic (e.g., Japanese hiragana, katakana)

o Alphabetic (e.g., Roman alphabet)

o Consonantal (e.g., Semitic scripts)

• These systems often evolved from early logo-syllabic systems, usually when cultures
interacted.

• The Arabic, Aramaic, Hebrew, Greek, and Roman alphabets all came from a West
Semitic script, likely adapted by Western Semitic mercenaries from a cursive form of
Egyptian hieroglyphs.

• Japanese syllabaries came from cursive Chinese characters used to represent sounds.

• Those Chinese characters had been used to phonetically represent Sanskrit in

Buddhist scriptures during China’s Tang dynasty.

4. Conceptual Foundation (Ur-theory)

• Sound-based writing implies the spoken word is made up of smaller units of

speech.
• This idea underlies modern phonology.
 • The decomposition of speech into smaller units is also the basis for:

o Speech recognition (turning acoustic waveforms into text)

o Speech synthesis / Text-to-speech (turning text into acoustic waveforms)

5. Focus of Chapter 7: Phonetics

• Phonetics studies linguistic sounds:

o Production by the articulators of the human vocal tract

o Acoustic realization of those sounds

o Digitization & processing of acoustic signals

6. Phones and Pronunciation in Technology

• Phones = individual speech units.

• Speech recognition systems need pronunciations for every word they can recognize.
• Text-to-speech systems need pronunciations for every word they can say.
• Phonetic alphabets are used to describe these pronunciations.

7. Two Main Areas of Phonetics

• Articulatory phonetics → how speech sounds are produced in the mouth.

• Acoustic phonetics → acoustic analysis of speech sounds.

8. Link to Phonology

• Phonology studies:

o How sounds vary systematically in different environments.

o How the sound system connects to the rest of grammar.

• Variation in pronunciation depending on context is crucial in speech modeling.

Important Points from the Passage

1. Whole language vs. phonics debate

o Modern reading-teaching debate.

o Mirrors a historical debate in writing systems.

2. Earliest writing systems

o Examples: Sumerian, Chinese, Mayan.

o Mostly logographic (one symbol represents a whole word).

 o Even early logographic systems included syllabic or phonemic elements
(symbols for sounds).

o Example: Sumerian symbol ba = “ration” and sound /ba/.

o Modern Chinese (still logographic) uses sound-based characters for foreign

words.

3. Purely sound-based writing systems

o Types:

▪ Syllabic (Japanese hiragana, katakana)

▪ Alphabetic (Roman alphabet)

▪ Consonantal (Semitic scripts)

o Origin: evolved from logo-syllabic systems.

o Often happened when two cultures met.

4. Historical origins

o Arabic, Aramaic, Hebrew, Greek, Roman alphabets came from West

Semitic script.

o Adapted by Western Semitic mercenaries from cursive Egyptian

hieroglyphs.

o Japanese syllabaries came from cursive Chinese characters used for sounds.

o Those Chinese characters were used to phonetically represent Sanskrit in

Buddhist scriptures during the Tang dynasty.

5. Ur-theory (basic idea behind phonology)

o Sound-based writing = spoken words are made of smaller speech units.

o Basis of modern phonology.

o Also basis for:

▪ Speech recognition → turning acoustic waveforms into text.

▪ Speech synthesis / Text-to-speech → turning text into acoustic

waveforms.

6. Phonetics (Chapter 7 introduction)

o Study of linguistic sounds:

▪ How produced by articulators in the human vocal tract.

▪ How acoustically realized.
 ▪ How digitized and processed.

7. Phones and pronunciation in technology

o Phones = individual speech units.

o Speech recognition: needs pronunciation for every recognizable word.

o Text-to-speech: needs pronunciation for every word it can speak.

o Phonetic alphabets are used to describe pronunciations.

8. Main areas of phonetics

o Articulatory phonetics → how speech sounds are produced in the mouth.

o Acoustic phonetics → acoustic analysis of speech sounds.

9. Phonology (link to phonetics)

o Studies:
▪ Systematic variation of sounds in different environments.
▪ Relation of sound system to the rest of grammar.

o Important for speech modeling because pronunciation changes with context.

Speech Sounds and Phonetic Transcription:

1. Phonetics

o Study of speech sounds used in the languages of the world.

o Pronunciation of a word is modeled as a string of symbols representing

phones or segments.

2. Phones

o A phone is a speech sound.

o Represented with phonetic symbols that may resemble letters in an alphabetic
language (like English).

3. Purpose of this section

o Surveys different phones in English, especially American English.

o Explains how they are produced and how they are represented
symbolically.

4. Phonetic Alphabets

o International Phonetic Alphabet (IPA)

▪ Developed in 1888 by the International Phonetic Association.
 ▪ Goal: transcribe the sounds of all human languages.

▪ Includes both an alphabet and principles for transcription.

▪ Same utterance can be transcribed in different ways according to IPA

principles.

o ARPAbet (Shoup, 1980)

▪ Designed specifically for American English.

▪ Uses ASCII symbols.

▪ Can be seen as an ASCII form of an American-English subset of IPA.

▪ Common in online pronunciation dictionaries and computational

applications where non-ASCII fonts are inconvenient.

5. Choice in this book

o Will use ARPAbet instead of IPA for computational purposes.

o Figures 7.1 (consonants) and 7.2 (vowels) show ARPAbet symbols with IPA
equivalents.

6. Rare Phones Example

o [ux]: Rare in General American English.

▪ Represents a fronted [uw] found in Western/Northern Cities dialects

from late 1970s.

▪ Popularized by imitations of “Valley Girls” (Moon Zappa, 1982).

▪ For most speakers, [uw] is still more common (e.g., dude [d uw d]).

7. ARPAbet Consonants – Examples

(With IPA equivalents)

o [p] → parsley [p aa r s l iy]

o [t] → tea [t iy]

o [k] → cook [k uh k]

o [b] → bay [b ey]

o [d] → dill [d ih l]

o [g] → garlic [g aa r l ix k]

o [m] → mint [m ih n t]

o [n] → nutmeg [n ah t m eh g]
o [ng] → baking [b ey k ix ng]
 o [f] → flour [f l aw axr]

o [v] → clove [k l ow v]

o [th] → thick [th ih k]

o [dh] → those [dh ow z]

o [s] → soup [s uw p]

o [z] → eggs [eh g z]

o [sh] → squash [s k w aa sh]

o [zh] → ambrosia [ae m b r ow zh ax]

o [ch] → cherry [ch eh r iy]

o [jh] → jar [jh aa r]

o [l] → licorice [l ih k axr ix sh]

o [w] → kiwi [k iy w iy]
o [r] → rice [r ay s]

o [y] → yellow [y eh l ow]

o [h] → honey [h ah n iy]

o Rare consonants:

▪ [q] → uh-oh [q ah q ow] (glottal stop)

▪ [dx] → butter [b ah dx axr] (flap)

▪ [nx] → winner [w ih nx axr] (nasal flap)

▪ [el] → table [t ey b el] (syllabic consonant)

8. ARPAbet Vowels – Examples

(With IPA equivalents)

o [iy] → lily [l ih l iy]

o [ih] → lily [l ih l iy]

o [ey] → daisy [d ey z iy]

o [eh] → pen [p eh n]

o [ae] → aster [ae s t axr]

o [aa] → poppy [p aa p iy]

o [ao] → orchid [ao r k ix d]
 o [uh] → wood [w uh d]

o [ow] → lotus [l ow dx ax s]

o [uw] → tulip [t uw l ix p]

o [ah] → buttercup [b ah dx axr k ah p]

o [er] → bird [b er d]

o [ay] → iris [ay r ix s]

o [aw] → sunflower [s ah n f l aw axr]

o [oy] → soil [s oy l]

o Reduced/Uncommon vowels:

▪ [ax] → lotus [l ow dx ax s] (schwa)

▪ [axr] → heather [h eh dh axr]

▪ [ix] → tulip [t uw l ix p] (reduced [ih])
▪ [ux] → dude [d ux d]

9. Orthography vs. Phonetic Symbols

o Many ARPAbet/IPA symbols match Roman letters (e.g., [p] in platypus).

o English spelling is opaque:

▪ Same letter can represent different sounds.

▪ Example: letter c → [k] in cougar [k uw g axr], but [s] in cell [s eh l].

o [k] can appear as:

▪ c (cougar)

▪ k (kangaroo)

▪ x (fox [f aa k s])

▪ ck (jackal [jh ae k el])

▪ cc (raccoon [r ae k uw n])

o Languages like Spanish have more transparent spelling-sound mapping than

English.

ARPAbet Consonants Table

ARPAbet IPA Example Word Example Transcription

p [p] parsley [p aa r s l iy]

ARPAbet IPA Example Word Example Transcription

t [t] tea [t iy]

k [k] cook [k uh k]

b [b] bay [b ey]

d [d] dill [d ih l]

g [g] garlic [g aa r l ix k]

m [m] mint [m ih n t]

n [n] nutmeg [n ah t m eh g]

ng [ŋ] baking [b ey k ix ng]

f [f] flour [f l aw axr]

v [v] clove [k l ow v]

th [θ] thick [th ih k]

dh [ð] those [dh ow z]

s [s] soup [s uw p]

z [z] eggs [eh g z]

sh [ʃ] squash [s k w aa sh]

zh [ʒ] ambrosia [ae m b r ow zh ax]

ch [tʃ] cherry [ch eh r iy]

jh [dʒ] jar [jh aa r]

l [l] licorice [l ih k axr ix sh]

w [w] kiwi [k iy w iy]

r [r] rice [r ay s]

y [j] yellow [y eh l ow]

h [h] honey [h ah n iy]

q [ʔ] uh-oh [q ah q ow]

ARPAbet IPA Example Word Example Transcription

dx [ɾ] butter [b ah dx axr]

nx [ɾ]̃ winner [w ih nx axr]

el [l̩ ] table [t ey b el]

ARPAbet Vowels Table

ARPAbet IPA Example Word Example Transcription

iy [i] lily [l ih l iy]

ih [ɪ] lily [l ih l iy]

ey [eɪ] daisy [d ey z iy]

eh [ɛ] pen [p eh n]

ae [æ] aster [ae s t axr]

aa [ɑ] poppy [p aa p iy]

ao [ɔ] orchid [ao r k ix d]

uh [ʊ] wood [w uh d]

ow [oʊ] lotus [l ow dx ax s]

uw [u] tulip [t uw l ix p]

ah [ʌ] buttercup [b ah dx axr k ah p]

er [ɝ] bird [b er d]

ay [aɪ] iris [ay r ix s]

aw [aʊ] sunflower [s ah n f l aw axr]

oy [ɔɪ] soil [s oy l]

ax [ə] lotus [l ow dx ax s]

axr [ɚ] heather [h eh dh axr]

ix [ɨ] tulip [t uw l ix p]
ARPAbet IPA Example Word Example Transcription

ux [ʉ] dude [d ux d]

ARTICULATORY PHONETICS

Study of how phones are produced as organs in the mouth, throat, and nose modify
airflow from the lungs.

1 The Vocal Organs

Airflow & Sound Production

• Air source: Rapid movement of air, usually expelled from the lungs.

• Air passes through:

1. Lungs →

2. Windpipe (Trachea) →

3. Larynx →

4. Mouth or Nose.

The Larynx & Vocal Folds

• Larynx = "Adam’s apple" / voice box.

• Contains vocal folds (vocal cords) → two small folds of muscle.

• Glottis = space between vocal folds.

Voicing

• Vocal folds close (but not tightly) → vibrate as air passes → Voiced sounds.

• Vocal folds far apart → no vibration → Unvoiced sounds.

Voiced Sounds Unvoiced Sounds

[b], [d], [g], [v], [z], all vowels [p], [t], [k], [f], [s], etc.

The Vocal Tract

• Area above trachea.

• Two parts:
 1. Oral tract (mouth)

2. Nasal tract (nose)

Nasal Sounds
• Air passes through nose (and also mouth cavity as resonator).

• English examples: m, n, ng.

Phones Classification

• Consonants

o Made by restricting/blocking airflow.

o Can be voiced or unvoiced.

o Examples: [p], [b], [t], [d], [k], [g], [f], [v], [s], [z], [r], [l].
• Vowels

o Minimal obstruction to airflow.

o Usually voiced, louder, and longer-lasting.

o Examples: [aa], [ae], [ao], [ih], [aw], [ow], [uw].

• Semivowels

o Share properties of both consonants and vowels.

o Voiced like vowels but short & less syllabic like consonants.
o Examples: [y], [w].

Diagram: The Vocal Organs

Nasal cavity

┌───────┐

│ │

Lips │ │ Teeth
↓ │ │

Tongue ────── Palate

│ │
 Pharynx │

↓ │

Larynx (vocal folds / glottis)

↓
Trachea (windpipe)

Lungs

ARTICULATORY PHONETICS

• Definition: Study of how phones (speech sounds) are produced as the mouth,
throat, and nose modify the airflow from the lungs.

• The ARPAbet phone list is meaningless without knowing how each phone is
produced.

7.2.1 The Vocal Organs

Airflow in Speech

• Speech is produced by rapid movement of air.

• Most sounds: Air from lungs → trachea (windpipe) → mouth or nose.

Larynx

• Known as Adam’s apple or voice box.

• Contains vocal folds (vocal cords) → 2 small muscle folds.

• Glottis: Space between vocal folds.

Voicing

• Vocal folds close (not tightly) → vibrate as air passes → Voiced sounds.

• Vocal folds far apart → no vibration → Unvoiced sounds.

Voiced Unvoiced

[b], [d], [g], [v], [z], all vowels [p], [t], [k], [f], [s], etc.

Vocal Tract
• Area above trachea.
 • Two parts:

1. Oral tract (mouth)

2. Nasal tract (nose)

Nasal Sounds

• Air passes through nose (also resonates in mouth).

• Examples: m, n, ng.

Phones Classification

1. Consonants

o Restrict/block airflow.
o May be voiced or unvoiced.
o Examples: [p], [b], [t], [d], [k], [g], [f], [v], [s], [z], [r], [l].

2. Vowels

o Minimal obstruction.

o Usually voiced, louder & longer-lasting.

o Examples: [aa], [ae], [ao], [ih], [aw], [ow], [uw].

3. Semivowels

o Properties of both vowels & consonants.

o Voiced like vowels, short like consonants.

o Examples: [y], [w].

Diagram – Vocal Organs (Side View)

(Shows airflow path and main parts)

scss

CopyEdit

Nasal Cavity

↑
┌─────────┐
│ │

Lips Teeth Alveolar Ridge

↓ ↓ ↓

Tongue ─── Palate ─── Velum (soft palate)

↓

Pharynx

Larynx (vocal folds, glottis)

Trachea

↓
Lungs

7.2.2 Consonants – Place of Articulation

• Definition: The point of maximum restriction of airflow when producing a

consonant.

• Used in automatic speech recognition to group phones.

Major English Places of Articulation

1. Labial

o Bilabial: Both lips come together.

Examples: [p] (possum), [b] (bear), [m] (marmot).

o Labiodental: Bottom lip touches upper teeth.

Examples: [v], [f].

2. Dental

o Tongue against teeth (tip slightly between).

Examples: [th] (thing), [dh] (though).

3. Alveolar

o Tip of tongue against alveolar ridge (behind upper teeth).

Examples: [s], [z], [t], [d].
o Coronal: Term for both dental + alveolar sounds.
 4. Palatal

o Palato-alveolar: Tongue blade against rising back of alveolar ridge.

Examples: [sh] (shrimp), [ch] (china), [zh] (Asian), [jh] (jar).

o Palatal proper: Front of tongue close to palate.

Example: [y] (yak).

5. Velar

o Back of tongue against velum (soft palate).

Examples: [k] (cuckoo), [g] (goose), [ng] (kingfisher).

6. Glottal

o Constriction at glottis (vocal folds close).

Example: Glottal stop [q] (IPA [ʔ]).

Diagram – Places of Articulation

scss
CopyEdit

[ Lips ] → Labial (Bilabial, Labiodental)

[ Teeth ] → Dental

[ Alveolar ridge ]→ Alveolar

[ Palate ] → Palatal

[ Velum ] → Velar

[ Glottis ] → Glottal

This has every term, example, and keyword exactly as in your text but in a clear point-
wise format with diagrams so it’s easy to study and reproduce in exams.
Consonants: Manner of Articulation:

Definition:

• Manner of articulation → How the restriction in airflow is made during consonant

production.

• Along with place of articulation, it usually identifies a consonant uniquely.

1. Stop (Plosive)
 • Airflow completely blocked for a short time → followed by an explosive release.

• Two phases:

o Closure → period of complete blockage.

o Release → explosion of air.

• Examples:

o Voiced stops: [b], [d], [g]

o Unvoiced stops: [p], [t], [k]

• Other notes:

o Also called plosives.

o Some systems distinguish closure and release separately (e.g., ARPAbet [pcl],
[tcl], [kcl]).

o Unreleased stops → no explosive release (e.g., end of words)

▪ ARPAbet: [pd], [td], [kd], [bd], [dd], [gd]

▪ IPA: [p̚], [t̚], [k̚]

o In this chapter → [p], [t], [k] = full stop with both closure & release.

2. Nasal

• Produced by lowering the velum → allows air to pass into the nasal cavity.

• Examples: [n], [m], [ŋ] (ng).

3. Fricatives
• Airflow constricted but not fully blocked → causes turbulent airflow.

• Produces a “hissing” sound.

• Examples by place:

o Labiodental: [f], [v] → lower lip against upper teeth.

o Dental: [θ] (“th” in thing), [ð] (“th” in though) → air flows around tongue
between teeth.

o Alveolar: [s], [z] → tongue against alveolar ridge, forcing air over teeth.

o Palato-alveolar: [ʃ] (“sh”), [ʒ] (“zh” as in Asian) → tongue at back of alveolar

ridge, air through tongue groove.
 • Sibilants → higher-pitched fricatives: [s], [z], [ʃ], [ʒ].

• Affricates → stop immediately followed by fricative: [tʃ] (“ch”), [dʒ] (“jh” as in

giraffe).

4. Approximants

• Articulators close together but not close enough for turbulence.

• Examples:
o [j] (“y” in yellow) → tongue near roof of mouth, no turbulence.

o [w] (“w” in wood) → back of tongue near velum.

o American [r] →

▪ Tongue tip near palate, OR

▪ Whole tongue bunched near palate.

o [l] → tip of tongue against alveolar ridge/teeth, sides lowered so air flows
over them.

▪ Called lateral sound (air passes along tongue sides).

5. Tap / Flap

• Quick motion of tongue against alveolar ridge.

• Example: [ɾ] → middle of lotus ([l ow dx ax s]) in American English.

• In many UK dialects → replaced by [t].

Diagram Needed

You should draw or insert:

1. Airflow blockage & release (for stops).

2. Velum lowered for nasals.

3. Constriction points for fricatives.

4. Positions for approximants & laterals.

5. Tap/flap quick contact motion.

Vowels:
Definition:
 • Like consonants, vowels are described by articulator positions during production.

• Three main parameters:

1. Vowel height → height of highest part of tongue.

2. Vowel frontness/backness → position of highest tongue point (toward front

or back).

3. Lip shape → rounded or unrounded.

1. Vowel Height

• High vowels → tongue raised high.

• Mid vowels → tongue in mid position.

• Low vowels → tongue lowered.

• Examples:

o High front: [iy] (heed)

o Low front: [ae] (had)

o High back: [uw] (who’d)

o [ih] higher than [eh] (both front vowels).

2. Vowel Frontness / Backness

• Front vowels → tongue highest point toward front of mouth.

o Examples: [iy], [ih], [eh], [ae]

• Back vowels → tongue highest point toward back of mouth.

o Examples: [uw], [uh], [ao], [aa]

3. Lip Shape

• Rounded vowels → lips rounded (like in whistling).

o Examples: [uw], [ao], [ow]

• Unrounded vowels → lips relaxed/spread.

4. Diphthongs
 • Definition: vowel where tongue position changes markedly during production.

• Represented in vowel charts as vectors instead of points.

• English is rich in diphthongs.

• Examples: [ay] (eye), [aw] (cow), [oy] (boy).

5. Important Notes

• Vowel height is schematic → correlates with acoustic patterns more than exact
tongue position.

• Figures to understand:

o Fig. 7.5 → tongue position examples for [iy], [ae], [uw].

o Fig. 7.6 → schematic vowel chart (high, mid, low; front–back positions;
monophthongs vs diphthongs).

If you want, I can make you a single combined diagram showing:

• Vowel chart (front–back, high–low)

• Examples of rounded/unrounded vowels

• Arrows for diphthongs

That way, your consonants + vowels notes will be visually complete.

UNIT 5
Speech synthesis
Speech Synthesis – Overview

Historical Background

• 1769, Vienna – Wolfgang von Kempelen built The Mechanical Turk for Empress
Maria Theresa.

o Appearance: Wooden box with gears + robot mannequin moving chess pieces
via mechanical arm.

o Toured Europe & Americas → defeated Napoleon Bonaparte, played

Charles Babbage.
o Hoax → secretly operated by a hidden human chess player.
 • 1769–1790 – Von Kempelen also built first full-sentence speech synthesizer (not a
hoax).

o Components:

▪ Bellows → simulated lungs.

▪ Rubber mouthpiece + nose aperture.

▪ Reed → simulated vocal folds.

▪ Whistles → fricatives.
▪ Auxiliary bellows → puff of air for plosives.

▪ Flexible leather “vocal tract” → adjusted to produce different

consonants & vowels.

o Operation: Controlled by moving levers with both hands, opening/closing

passages.

Modern Speech Synthesis (Text-to-Speech, TTS)

• Definition: Generating speech (acoustic waveforms) from text input.

• Applications:

1. Conversational agents – telephone-based dialogue systems (with speech

recognition).

2. Non-conversational systems – reading aloud for the blind, video games, toys.

3. Assistive communication – e.g., Stephen Hawking (ALS) typed words →

synthesizer speech output.

• Limitations:

o Even best systems may sound wooden.

o Limited voice variety.

• State of the art:

o Can produce remarkably natural speech for wide range of inputs.

Basic TTS Process

• Text Analysis – Convert text into phonemic internal representation.

 o Includes: Expanding acronyms (e.g., PG&E → “P G AND E”), converting
numbers (20 → “twentieth”), assigning phone sequences, prosody, and
phrasing.

• Waveform Synthesis – Convert internal representation into speech waveform.

Example (Fig. 8.1)

Sentence: PG&E will file schedules on April 20.

• Converted to:

o Expanded words: P G AND E WILL FILE SCHEDULES ON APRIL

TWENTIETH

o Phones: p iy jh iy ae n d iy w ih l f ay l s k eh jh ax l z aa n ey p r ih l t w eh n
t iy ax th

o Prosodic markers: * * * L-L%

Main Waveform Synthesis Paradigms

1. Concatenative synthesis (focus of chapter) –

o Store recorded speech samples in database.

o Chop and recombine to form new sentences.

o Unit selection synthesis → chooses best matching units from large database.

2. Formant synthesis –

o Model resonance characteristics of vocal tract.

o Fully artificial, not sample-based.

3. Articulatory synthesis –
o Simulate physical processes of speech production.

Architecture Example

• Hourglass metaphor (Taylor, 2008) → Two-step narrowing from text to phones, then
expanding from phones to waveform.

• Modern commercial TTS → Mostly based on concatenative unit selection synthesis.

If you want, I can also make this into a clear flowchart showing:
Historical → Modern TTS → Text Analysis → Waveform Synthesis → Methods
(Concatenative, Formant, Articulatory)

Text Normalization:

Purpose

• Goal: Convert raw text into a phonemic internal representation for speech
synthesis.

• Why needed: Raw text contains abbreviations, numbers, punctuation quirks, and
irregular formats that must be standardized for correct pronunciation.

Main Steps in Text Normalization

1. Sentence Tokenization

• Definition: Segmenting text into separate sentences/utterances for synthesis.

• Challenges:
o Abbreviations with periods – Avoid splitting sentences incorrectly.

▪ Example: B.C. Hydro → do not treat the period after B.C. as an end of
sentence.

o Non-standard sentence boundaries – Detect sentences ending with

punctuation other than a period.
▪ Example: Sentence ending after “collected” even though it uses a
colon instead of a period.

2. Handling Non-Standard Words (NSWs)

• Definition: Words/symbols that require expansion into standard spoken form.

• Types & Examples:

o Dates:

▪ March 31 → “March thirty-first” (not “March three one”).

o Numbers:

▪ $1 billion → “one billion dollars” (insert word dollars after “billion”).

o Acronyms:
▪ Expand into letter-by-letter or full-word form as appropriate (e.g.,
PG&E → “P G and E”).
 o Abbreviations:

▪ Must identify correct expansions (e.g., Dr. → “Doctor”, St. → “Street”

or “Saint” depending on context).

Example from Enron Corpus (Klimt & Yang, 2004)

• Raw text contains:

o Abbreviation with period (B.C. Hydro).

o Date (March 31).

o Currency & number ($1 billion).

o Colon as sentence end (“collected:” → signals end of an utterance).

Key Output of Text Normalization

• Clean, sentence-segmented text.

• All non-standard words expanded into their full spoken forms.

• Ready for phonetic analysis and prosodic analysis in the TTS pipeline.

8.1.1 Sentence Tokenization

Definition

• Sentence Tokenization = Process of detecting sentence boundaries in text.

• Goal: Correctly identify where sentences begin and end for speech synthesis.

Challenges in Sentence Tokenization

1. Punctuation Ambiguity

• Periods are not always sentence boundaries.

• Examples:

o Abbreviations – B.C. Hydro (period is part of abbreviation, not sentence end).

o Dual Role – When abbreviation ends a sentence (Dr. J. M. Freeman).

o Non-period boundaries – Colons can signal sentence end:

▪ collected: “We continue…”

o Example references: (8.2), (8.3), (8.4).
2. Period Disambiguation

• Definition: Deciding whether a period marks End-of-Sentence (EOS) or not.

• Approach:
o Early method: Simple Perl scripts (Ch. 3).

o Modern method: Machine Learning–based EOS classifier.

Machine Learning Approach for Sentence Tokenization

Training Process

1. Hand-label a training set with correct sentence boundaries.

2. Tokenize text into tokens separated by whitespace.

3. Select candidate tokens containing potential boundary punctuation:
o ., !, ? (possibly also :).

4. Train a classifier to predict EOS vs not-EOS for each candidate.

Features for EOS Classification

A. Basic Feature Templates

• Prefix: Text before punctuation in the token.

• Suffix: Text after punctuation in the token.

• Abbreviation Check: Whether prefix/suffix is in abbreviation list.

• Previous Word and Next Word.

• Whether previous word is abbreviation.

• Whether next word is abbreviation.

Example (8.5): ANLP Corp. chairman Dr. Smith resigned.

• Candidate = period in "Corp."

o PreviousWord = ANLP
o NextWord = chairman

o Prefix = Corp
o Suffix = NULL
 o PreviousWordAbbreviation = 1

o NextWordAbbreviation = 0

B. Lexical Probability Features

• Probability that the candidate token occurs at end of sentence.

• Probability that the word after candidate occurs at beginning of sentence.

C. Language-Specific Features

• Capitalization patterns:

o Case of candidate word: Upper, Lower, AllCap, Numbers.

o Case of following word: Upper, Lower, AllCap, Numbers.

• Special abbreviation classes:
o Honorifics/Titles – Dr., Mr., Gen.

o Corporate designators – Corp., Inc.

o Month abbreviations – Jan., Feb.

Classification Methods

• Common algorithms: Logistic Regression, Decision Trees.

• Logistic Regression → Often higher accuracy.

• Decision Trees → More interpretable (example shown in Fig. 8.3).

• Features used:

o Log likelihood of current word being sentence start (bprob).

o Log likelihood of previous word being sentence end (eprob).

o Capitalization of next word.

o Abbreviation subclass (company, state, unit of measurement).

Non-Standard Words in Text Normalization :

Definition

Non-standard words (NSWs) are tokens like numbers or abbreviations that must be expanded
into full English words before pronunciation.
Ambiguity in NSWs

• Numbers can have multiple pronunciations depending on context:

• 1750 → seventeen fifty (year)

• 1750 → one seven five zero (password)
• 1750 dollars → seventeen hundred and fifty OR one thousand seven hundred and fifty
• Roman numerals (e.g., IV) can mean four, fourth, or letters "I V" (intravenous).
• Fractions like 2/3 → two thirds, February third, or two slash three.

Types of NSWs

• Numbers

• NUM: number (cardinal) – 12, 45, 1/2, 0.6

• NORD: ordinal – 3rd, May 7
• NDIG: number as digits – Room 101
• NIDE: identifier – 747, I5
• NADDR: street address – 386 Main St.
• NZIP: zip code – 91020
• NTIME: time – 3.20, 11:45
• NDATE: date – 2/28/05
• NYER: years – 1998, 80s, 2008
• MONEY: $3.45, Y20,200
• BMONEY: billions/trillions – $3.2 billion
• PRCT: percentage – 75%

• Alphabetic

• EXPN: abbreviations – N.Y., mph, gov’t

• LSEQ: letter sequences – DVD, IBM
• ASWD: acronyms as words – NASA, IKEA

• Realization Patterns

• Paired method (e.g., years) → seventeen fifty for 1750.

• Serial method (e.g., zip codes) → nine four one one zero.
• BMONEY → always read with currency word at the end.

• Processing NSWs

Step 1: Tokenization

• Tokenize by whitespace.

• Identify tokens not in the pronunciation dictionary as NSWs.

Handle:
 • Known abbreviations in dictionaries (e.g., st, mr, mrs, mon, tues, nov).
• Single-character tokens.
• Hyphenated words (2-car).
• CamelCase words (RVing).

Step 2: Classification

• Assign NSW type (NUM, LSEQ, EXPN, etc.) using:

• Regular expressions (e.g., NYER: /1[89][0-9][0-9]|20[0-9][0-9]/).
• Machine learning classifiers with features:

i Letter patterns (all caps, two vowels, contains slash, token length).
ii Context words (Chapter, on, king).
iii Neighboring word identity.

Step 3: Expansion

o EXPN: needs abbreviation dictionary + homonym disambiguation.

• LSEQ: expand letter by letter.
• ASWD: keep as is.
• NUM/NORD: expand to cardinal/ordinal words.
• NDIG/NZIP: digit-by-digit expansion.
• NYER: two-digit pairs (nineteen eighty), except:

➢ Ends in "00" → read as cardinal (two thousand).

➢ Hundreds method → eighteen hundred.

• NTEL: digit sequence, or paired digits for last four, or trailing unit method (five
thousand for last digits).

Dialect & Language-Specific Variations

• Australian English → double three for "33".

• French: depends on gender of noun (un garçon, une fille).
• German: morphological case affects pronunciation (Heinrich IV changes form).

Homograph Disambiguation :

Goal:

• To determine correct pronunciation for homographs — words with the same spelling
but different pronunciations.

Definition & Examples

Homographs: Same spelling, different pronunciations.

• English examples:
 1. use

• noun: /y uw s/ → It’s no use

• verb: /y uw z/ → to use the telephone

2. live
• verb: /l ih v/ → Do you live near a zoo
• adjective: /l ay v/ → live animals

• bass

• fish: /b ae s/ → bass fishing

• instrument: /b ey s/ → bass guitar

French examples:

• fils: [fis] ‘son’ vs [fil] ‘thread’

• fier: ‘proud’ vs ‘to trust’
• est: ‘is’ vs ‘East’

Part-of-Speech (POS) as a Disambiguation Tool

In English (and similar languages like French, German), different homograph forms often
differ in part-of-speech.

Example:

• use: noun vs verb

• live: verb vs adjective

Fig. 8.5 Patterns:

o Final voicing: noun: /s/ vs verb: /z/ → use, close, house

o Stress shift: noun: initial stress vs verb: final stress → record, insult, object

o -ate final vowel weakening

• noun/adjective: final /ax/ vs verb: final /ey/ → estimate, separate, moderate

Statistical Evidence

Liberman & Church (1992):

• Many frequent homographs in AP newswire corpus can be disambiguated via POS.

• Top 15 most frequent homographs: use, increase, close, record, house, contract, lead,
live, lives, protest, survey, project, separate, present, read

Standard Disambiguation Approach

 Store distinct pronunciations for homographs labeled by POS.

• Run a POS tagger to select the correct pronunciation in context.

When POS Alone Is Not Enough

o Same POS, different pronunciations:

• bass: fish (/b ae s/) vs instrument (/b ey s/)
• lead: noun (/l iy d/) leash vs noun (/l eh d/) metal

Abbreviation ambiguity:

• Dr. → doctor vs drive

• St. → Saint vs street

Capitalization differences:

• polish vs Polish (homographs in sentence-initial or all-caps text)

Practical TTS Handling

o Many of these harder homographs are ignored in TTS systems.

• Alternatively, use Word Sense Disambiguation (WSD) methods:
• Example: Decision-list algorithm (Yarowsky, 1997)

Phonetic Analysis:
Purpose:
• Take normalized word strings from text analysis and generate pronunciations for each
word.

Main Component

• Large pronunciation dictionary (core tool for phonetic analysis).

Why Dictionaries Alone Are Not Enough

• Running text often contains words not in the dictionary.

• Example (Black et al., 1998):

o British English dictionary: OALD lexicon tested on first section of Penn Wall
Street Journal Treebank.
o Total words (tokens): 39,923
 o Not in dictionary: 1,775 tokens (4.6%)

▪ Unique types: 943

o Distribution of unseen word tokens:

Category Count %

Names 1,360 76.6%

Unknown words 351 19.8%

Typos & others 64 3.6%

• Two main areas needing augmentation:

1. Handling names

2. Handling other unknown words

Process Order

1. Dictionaries

2. Names

3. Grapheme-to-phoneme rules for other unknown words

8.2.1 Dictionary Lookup

Phonetic Dictionaries

• Introduced earlier in Ch. 8.

• Commonly used in TTS:

o CMU Pronouncing Dictionary (CMUdict, 1993)

▪ ~120,000 words

▪ Pronunciations: roughly phonemic from 39-phone ARPAbet-derived

set

▪ Phonemic transcription:
▪ No surface reductions (like [ax], [ix]) directly written

▪ Each vowel has stress tags:

▪ 0 → unstressed
▪ 1 → stressed
 ▪ 2 → secondary stress

▪ Non-diphthong vowels with stress 0 usually → [ax] or [ix]

▪ Most words have one pronunciation, ~8,000 words have two or three

▪ Some phonetic reductions are reflected

▪ Not syllabified (nucleus implicitly marked by numbered vowel)

▪ Limitations:

▪ Designed for speech recognition, not synthesis

▪ Does not specify which pronunciation to choose for synthesis

▪ No syllable boundaries

▪ Capitalization of headwords → can’t distinguish US vs us

▪ US: [AH1 S] or [Y UW1 EH1 S]

Sample CMUdict Entries (Fig. 8.6)
ANTECEDENTS AE2 N T IH0 S IY1 D AH0 N T S

PAKISTANI P AE2 K IH0 S T AE1 N IY0

CHANG CH AE1 NG

TABLE T EY1 B AH0 L

DICTIONARY D IH1 K SH AH0 N EH2 R IY0

TROTSKY T R AA1 T S K IY2

DINNER D IH1 N ER0

WALTER W AO1 L T ER0

LUNCH L AH1 N CH

WALTZING W AO1 L T S IH0 NG

MCFARLAND M AH0 K F AA1 R L AH0 N D

WALTZING(2) W AO1 L S IH0 NG

UNISYN Dictionary

• ~110,000 words
• Freely available for research

• Designed for speech synthesis (Fitt, 2002)

• Features:
 o Syllabifications

o Stress markings

o Some morphological boundaries

o Can produce pronunciations in multiple English dialects:

▪ General American

▪ RP British

▪ Australian English

▪ …and dozens more

o Slightly different phone set than CMUdict

UNISYN Examples:
going: { g * ou }.> i ng >
antecedents: { * a n . tˆ i . s ˜ ii . d n! t }> s >

dictionary: { d * i k . sh @ . n ˜ e . r ii }

8.2.2 Names

Importance of Names in Speech Synthesis

• Names are a major source of pronunciation errors in TTS.

• Categories of names:

1. Personal names – first names and surnames

2. Geographical names – cities, streets, and other place names

3. Commercial names – company and product names

Scale of the Problem

• Personal names in the U.S. (Spiegel, 2003, based on Donnelly and household lists):

o ~2 million different surnames

o ~100,000 first names

• Comparison:

o 2 million surnames is an order of magnitude larger than the entire CMU

Pronouncing Dictionary (~120,000 words).
Large-Scale TTS Solutions

• Most large-scale TTS systems include a large name pronunciation dictionary.

• CMU Pronouncing Dictionary:

o Contains many names

o Includes:

▪ The most frequent 50,000 surnames (from Bell Labs estimate of U.S.
personal name frequency)

▪ ~6,000 first names

Coverage Studies

• Liberman & Church (1992):

o A dictionary of 50,000 names covered 70% of name tokens in a 44-million-

word AP newswire corpus.

o Many remaining names could be derived from these 50,000:

▪ Up to 97.43% coverage possible by applying simple modifications to

the known names.

Name Formation Techniques

1. Adding stress-neutral suffixes to known names

o Examples:

▪ walters = walter + s

▪ lucasville = lucas + ville

▪ abelson = abel + son

2. Rhyme analogy

o Example:
▪ Known: Trotsky → /tr/ replaced with /pl/ → Plotsky

3. Morphological decomposition

4. Analogical formation
5. Mapping unseen names to spelling variants already in the dictionary
 o (Fackrell & Skut, 2004)

Current Challenges

• Name pronunciation remains difficult despite these techniques.

• Many modern systems handle unknown names via grapheme-to-phoneme (G2P)
methods:

o Often build two separate predictive systems:

▪ One for names

▪ One for non-names

Additional Resources

• Spiegel (2003, 2002) provides further details on issues in proper name

pronunciation.

8.2.3 Grapheme-to-Phoneme (G2P):

Definition

• Grapheme-to-Phoneme Conversion (G2P):

Converting a sequence of letters (graphemes) into a sequence of phones (phonemes).

o Example: "cake" → [K EY K]

• Also called Letter-to-Sound (LTS) conversion.

Historical Approach – Hand-Written Rules

• Based on Chomsky–Halle phonological rewrite rules (Chapter 7 format).

• Rules are applied in order:

o Earlier rules = context-specific

o Later rules = default (apply only if earlier ones don’t)

• Example simple rules for letter c:

1. c → [k] / {a, o} V (context-dependent)

2. c → [s] (default, context-independent)

• Actual rules are more complex:

o c → [ch] in cello, concerto
 • Stress rules for English are especially complex:

o Example rule from Allen et al. (1987):

Assign primary stress in specific contexts involving weak syllables and
morpheme-final positions:

1. Before a weak syllable + morpheme-final short vowel + consonants

(e.g., difficult)

2. Before a weak syllable + morpheme-final vowel (e.g., oregano)

Modern Approach – Probabilistic Methods

• First formalized by Lucassen & Mercer (1984):

o Goal: Find the most probable phone sequence P for a letter sequence L:

P^=arg⁡max⁡PP(P∣L)\hat{P} = \arg\max_P P(P|L)P^=argPmaxP(P∣L)

• Requires:

o Training set: Words with spelling + pronunciation

o Test set: New words to predict

Step 1 – Letter-to-Phone Alignment (Training Set)

• Needed: Which phones align with which letters.

• Some letters align to:

o Multiple phones: x → [k s]

o No phones: final e in cake → ε (null)

• Example alignment:

makefile
CopyEdit

L: c a k e

| | | |

P: K EY K ε

Semi-Automatic Alignment Method (Black et al., 1998)

• Relies on allowable phone lists for each letter.

• Example:
 o c: k, ch, s, sh, t-s, ε

o e: ih, iy, er, ax, ah, eh, ey, uw, ay, ow, y-uw, oy, aa, ε

• Process:

1. For each word, find all alignments matching allowable lists.

2. Count (letter, phone) pairs across all alignments.

3. Compute probability:

P(pi∣lj)=count(pi,lj)count(lj)P(p_i | l_j) = \frac{\text{count}(p_i, l_j)}{\text{count}(l_j)}P(pi

∣lj)=count(lj)count(pi,lj)

4. Use Viterbi algorithm to find best alignment A for each (P, L) pair.

Step 2 – Predicting Phones for New Words (Test Set)

• Train machine learning classifier (e.g., decision tree) on aligned training set.

• Predicts most probable phone for each letter.

Features for Decision Tree

1. Window of surrounding letters:

o k previous + k following letters

o Example:

▪ "cat": a → [AE]

▪ "cake": a → [EY] (influenced by final e)

2. Previous predicted phone:

o Gives phonotactic context

o Must process left-to-right to use previous output.

3. Stress prediction:
o Augment vowels with stress info:

▪ 2 levels: AE / AE1

▪ 3 levels: AE0, AE1, AE2 (as in CMU lexicon)

4. Part-of-speech tag of the word:

o Even for unknown words (estimated by POS tagger)

5. Previous vowel stress status
 6. Letter classes:

o Consonants, vowels, liquids, etc.

7. Following word features (language-specific phenomena):

o Example: French liaison:

▪ six → [sis] (j’en veux six)

▪ [siz] (six enfants)

▪ [si] (six filles)

Special Handling for Names

• Many systems build two separate G2P models:

1. For unknown personal names

2. For other unknown words
• Additional features for names:

o Foreign language origin (predicted via letter sequence n-gram models)

Model Type

• Decision tree = conditional classifier:

o Finds phone sequence with highest conditional probability given grapheme

sequence.

• More recent models:

o Joint classifier:
▪ Hidden state = graphone (grapheme + phoneme pair)

This is now a full extraction—nothing skipped, all examples, rules, and equations preserved,
plus structured for clarity.