HG4052 / HG7034 · Poster Fair
HG4052 / HG7034 · Speech Synthesis & Recognition / Week 13

Poster
Fair.

A mini conference on synthetic voices.

Fri 13 Nov 2026 · TR69 21 poster pitches Career Panel: Ezio Xie

Chenzi Xu · chenzi.xu@ntu.edu.sg

You paid for the kopi?
★
The program · Fri 13 Nov 2026 · TR69

Today is a mini conference

09:40
Warm-up
The table trick from Quiz 2
One more look at how the table finds the best path
a few minutes
09:45
Career Panel
My journey through linguistics
Ezio Xie (Founder, PDP Limited): panel sharing + Q&A
45–60 min
about10:45
Poster pitches
21 posters, 3 minutes each
Quick rotation: one voice after another
about 75 min
about12:00
Closing reception
Food and chat
Chat with the presenters
informal
?
Quiz 2, B1 · the question

One table, three rules

The word “see”: two states, and what each arrow costs
Emission probabilities: how well each frame fits each state
ON THE PAPER

(a)(i) Fill in the Viterbi table. The word ends after frame 3, so do not count the transition out of /i/. [2]

(a)(ii) Give the best path and its probability. [1]

(b) When several routes arrive at the same cell, what do Viterbi and the CTC forward table put there, and why? [2]

(c)(i) Forced alignment fills the same kind of table, with the transcript already known. What does it already know, and what does it give you? [1]

(d) All three methods fill a table. What is the benefit of that? [1]

↘
Quiz 2, B1(c) and (d) · the same table in forced alignment

Forced alignment: the path only steps down

The word “kopi”: four phones, ten frames
Made-up probabilities. Simplified: one state per phone; real aligners use about three per phone and allow optional pauses between words.
  • Rows: the transcript’s phones, in order
  • Each frame: stay, or step down one
  • Column 1 is 0 except the first phone
  • What we find: the boundaries
B1(d) · WHAT THE TABLE SAVES

THE QUIZ WAS ONE

V
Quiz 2, B1(a)(i) · one cell at a time

Each cell keeps the best route into it

V(s, t) = P(xt | s) × max over routes of [ V(prev, t−1) × P(prev → s) ]

V(s, t): the probability of the best path that ends in state s at frame t

a route: one cell one frame back, times the transition out of it

max: keep the biggest route, remember where it came from, discard the rest

Colors, in every figure today

● /s/   ● /i/   ● the route kept, and the best path   ● the route discarded

×
Quiz 2, B1(a)(i) · the last column

Compare the routes, not the cells

Two routes arrive at /i/ in frame 3
  • 0.18 is the larger cell, and it loses: its transition is the smaller one
  • A route is a cell times the transition out of it, so both parts count
  • Keep 0.06, and remember that it came from /s/
WHAT THIS COSTS

Comparing the cells instead of the routes gives 0.0216 and the path /s/ → /i/ → /i/. It was the most common wrong answer, and it scored nothing on (a)(ii).

←
Quiz 2, B1(a)(ii) · reading the answer off

Follow the kept routes, not the maxima

  • Start at the cell the word has to end in, then follow the route each cell kept
  • A cell holds the best path into it, and those paths need not agree with each other
WHY 0.008 AND NOT 0

V(/s/, 3) scores the best route into that cell, and a cell knows nothing about what follows it. The end-in-/i/ rule applies when you read the answer off, not while you fill the table: with a fourth frame, /s/ is live again.

→
A question from the room

A word model only runs forward

The word model

An arc from /i/ back to /s/ would mean the pronunciation went /s i s/, which is a different word. The rows also leave no room for it: 0.4 + 0.6 and 0.3 + 0.7 already come to 1.

The ice-cream model

Hot and cold can follow each other in any order, so every arc exists and the matrix is full. Nothing fixes the sequence in advance.

What is still shared

The same recursion, the same table, the same backtrace. Only the set of allowed arcs differs, and that is what the word's spelling buys you.

1
Quiz 2, B1 · what the table buys you

One table does the work of every path

WHY THE REUSE IS SAFE

Nothing is approximated: what comes next depends only on the state a path is in, not on the way in, so a dropped route could never have won later.

2
Quiz 2, B1(b) to B1(d) · the same table, four uses

Keep the best route, or add them all

Viterbi decoding(the quiz)
Forced alignment
Unit selection
CTC forward
Known in advance
the frames
the frames and the phones
the units the text needs
the frames and the transcript
What is searched
the state sequence
the timing
a recording for each unit
no single path: all alignments
Columns
frames
frames
target units
frames
Rows
states
phones, in order
candidate recordings
tokens, with blanks
Allowed moves
the model’s arcs
stay, or step down one
any row to any row
stay, step down, skip a blank
In each cell
multiply, keep the largest
multiply, keep the largest
add costs, keep the smallest
multiply, add them all
Read off
the path
the boundaries
the recordings to splice
the transcript’s probability
cost = −log probability

Adding costs and keeping the smallest picks the same route as multiplying probabilities and keeping the largest. Unit selection keeps the best route like Viterbi; it just counts in costs, and any recording can follow any other.

Career Panel · 09:45

My journey through linguistics.

Panel sharing + Q&A

Ezio Xie
Ezio Xie
Founder, PDP Limited · 45–60 minutes
Career Panel · meet our guest
Ezio Xie, smiling Our guest!

Ezio Xie.

Founder, PDP Limited
Researcher in Linguistics and Computation
Education
MPhil in Linguistics
University of Oxford
?
Career Panel · Q&A

Ask Ezio anything

Question starters, if you need one:

What from your linguistics degree do you use most often?
You speak many languages. What motivates you to learn more?
What challenges did you face moving from linguistics into tech?
What does a typical week look like for you now?
Next · about 10:45

Poster pitches.

21 posters · 3 minutes each · quick rotation

Poster pitches · how it runs

Rules for the pitch

1
3 minutes
Hard stop at the chime
2
Poster up
Audio ready to play
3
Next up?
Be ready to go
4
Applause
For every voice
Tonight, 23:59: abstract and poster on NTULearn
3:00
Poster pitches · running order

The pitch clock

3:00
ready
Now Poster 1 1 of 21
Next Poster 2
Enter start or pause · N next pitch · B back · R reset · click a name to jump
Poster pitches · your vote

Vote: best pitch and best poster

Best pitch
The three minutes you remember
Best poster
The poster that taught you most
A number and a short reason for each · not your own
Closing · the course in one slide

Speech technology keeps coming back to linguistics

Rules
1950s–1980s
Formant synthesis, hand-written rules
Linguists write the rules
Statistics
1990s–2000s
HMMs, n-grams, unit selection
Phone sets, lexicons, labels
Neural
2010s
CTC, attention, neural vocoders
Data replaces the rules
Speech LLMs
2020s
Codec models, few-second cloning
Fluent, but whose speech?
Next
now
Linguistics of and with speech models
Linguists test, explain, repair
What you showed today
Minimal pairs, not scores
Plain against Singapore English, everything else fixed
A variety, not an error
Particles flattened to a full stop; local words and names anglicized
Ears, not just metrics
WER rose on local words while predicted naturalness stayed flat; clones sounded like the speaker without speaking like them
Closing · today’s posters as a research agenda

Six directions for the linguistics of speech technology

Language varieties and representational fairness
Are varieties modeled as systems in their own right, or as noise around a standard?
Seen today: 2, 8, 9, 10, 11, 17
Prosodic and pragmatic competence
Can a system map meaning onto intonation, focus and discourse particles?
Seen today: 12, 13, 14, 15, 16, 18
Diagnostic evaluation
Minimal pairs and expert listening that explain a capability, not a score
Seen today: 3, 5, 21, 26
Speaker identity and consent
What a voice clone preserves, what it loses, and who may authorize it
Seen today: 19, 20, 22, 27
Multilingual and code-switched speech
Loanwords, mixed-language utterances and low-resource languages
Seen today: 23, 25
Linguistically informed interpretability
Where are phones, tones and stress represented inside the network?
Open: no poster yet
Numbers are today’s pitches. Framing inspired by an EMNLP 2026 keynote on the linguistics of LLMs.
HG4052 / HG7034 · the last session

Thank
you.

To Ezio, to all 21 presenters, and to all of you.

Closing reception: food and chat, now

Tonight, by 23:59: abstract and poster on NTULearn

Makan time lah!