Pronunciation Modeling Lecture 11 Spoken Language Processing Prof. Andrew Rosenberg.

Pronunciation Modeling

Lecture 11

Spoken Language Processing

Prof. Andrew Rosenberg

2

What is a pronunciation model?

Acoustic Model

PronunciationModel

LanguageModel

Audio Features

Phone Hypothese

Word Hypothese

Word Hypothese

3

Why do we need one?

• The pronunciation model defines the mapping between sequences of phones and words.

• The acoustic model can deliver a one-best, hypothesis – “best guess”.

• From this single guess, converting to words can be done with dynamic programming alignment.

• Or viewed as a Finite State Automata.

4

Simplest Pronunciation “model”

• A dictionary.• Associate a word (lexical item,

orthographic form) with a pronunciation.

ACHE EY KACHES EY K SADJUNCT AE JH AH NG K TADJUNCTS AE JH AN NG K T SADVANTAGE AH D V AE N T IH JHADVANTAGE AH D V AE N IH JHADVANTAGE AH D V AE N T AH JH

5

Example of a pronunciation dictionary

6

Finite State Automata view

• Each word is an automata over phones

EY K

EY K

AH D V AE N T

S

I JH

7

Size of whole word models

• these models get very big, very quickly

EY K

EY K

AH D V AE N T

S

I JH

START END

8

Potential problems

• Every word in the training material and test vocabulary must be in the dictionary

• The dictionary is generally written by hand• Prone to errors and inconsistencies

ACHE EY KACHES EY K SADJUNCT AE JH AH NG K TADJUNCTS AE JH AN NG K T SADVANTAGE AH D V AE N T IH JHADVANTAGE AH D V AE N IH JHADVANTAGE AH D V AE N T AH JH

9

Baseforms represented by graphs

10

Composition

• From the word graph, we can replace each phone by its markov model

11

Automating the construction

• Do we need to write a rule for every word?

• pluralizing?– Where is it +[Z]? +[IH Z]?

• prefixes, unhappy, etc.– +[UH N]– How can you tell the difference between

“unhappy”, “unintelligent” and “under” and “

12

Is every pronunciation equally likely?

• Different phonetic realizations can be weighted.

• The FSA view of the pronunciation model makes this easy.

ACAPULCO AE K AX P AH L K OWACAPULCO AA K AX P UH K OWTHE TH IYTHE TH AXPROBABLY P R AA B AX B L IYPROBABLY P R AA B L IYPROBABLY P R AA L IY

13

Is every pronunciation equally likely?

• Different phonetic realizations can be weighted.

• The FSA view of the pronunciation model makes this easy.

ACAPULCO AE K AX P AH L K OW0.75ACAPULCO AA K AX P UH K OW

0.25THE TH IY

0.15THE TH AX

0.85PROBABLY P R AA B AX B L IY

0.5PROBABLY P R AA B L IY

0.4PROBABLY P R AA L IY

0.1

14

Collecting pronunciations

• Collect a lot of data• Ask a phonetician to phonetically

transcribe the data.• Count how many times each

production is observed.

• This is very expensive – time consuming, finding linguists.

15

Collecting pronunciations

• Start with equal likelihoods of all pronunciations

• Run the recognizer on transcribed speech– forced alignment

• See how many times the recognizer uses each pronunciation.

• Much cheaper, but less reliable

16

Out of Vocabulary Words

• A major problem for Dictionary based pronunciation is out of vocabulary terms.

• If you’ve never seen a name, or new word, how do you know how to pronounce it?– Person names– Organization and Company Names– New words “truthiness”, “hypermiling”,

“woot”, “app”– Medical, scientific and technical terms

17

Collecting Pronunciations from the web

• Newspapers, blog posts etc. often use new names and unknown terms.

• For example:– Flickeur (pronounced like Voyeur) randomly

retrieves images from Flickr.com and creates an infinite film with a style that can vary between stream-of-consciousness, documentary or video clip.

– Our group traveled to Peterborough (pronounced like “Pita-borough”)...

• The web can be mined for pronunciations [Riley, Jansche, Ramabhadran 2009]

18

Grapheme to Phoneme Conversion

• Given a new word, how do you pronounce it.

• Grapheme is a language independent term for things like “letters”, “characters”, “kanji”, etc.

• With a phoneme to grapheme-to-phoneme converter, dictionaries can be augmented with any word.

• Some languages are more ambiguous than others.

19

Grapheme to Phoneme conversion

• Goal: Learn an alignment between graphemes (letters) and phonemes (sounds)

• Find the lowest cost alignment.• Weight rules, and learn contextual variants.

T E X - T

T EH K S T

T E X T - - - - -

- - - - T EH K S T

20

Grapheme to Phoneme Difficulties

• How to deal with Abbreviations– US CENSUS– NASA, scuba vs. AT&T, ASR– LOL– IEEE

• What about misspellings?– should “teh” have an entry in the dictionary?– If we’re collecting new terms from the web,

or other unreliable sources, how do we know what is a new word?

21

Application of Grapheme to Phoneme Conversion

• This Pronunciation Model is used much more often in Speech Synthesis than Speech Recognition

• In Speech Recognition we’re trying to do Phoneme-to-Grapheme conversion– This is a very tricky problem.– “ghoti” -> F IH SH– “ghoti” -> silence

22

Approaches to Grapheme to Phoneme conversion

• “Instance Based Learning”– Lookup based on a sliding window of 3

letters– Helps with sounds like “ch” and “sh”

• Hidden Markov Model– Observations are phones– States are letters

23

Machine Learning for Grapheme to Phoneme Conversion

• Input:– A letter, and surrounding context, e.g. 2

previous and 2 following letters

• Output:– Phoneme

24

Decision Trees

• Decision trees are intuitive classifiers– Classifier: supervised machine

learning, generating categorical predictions

Feature > threshold?

Class A Class B

25

Decision Trees Example

26

Decision Tree Training

• How does the letter “p” sound?• Training data

– P loophole, peanuts, pay, apple– F physics, telephone, graph, photo– ø apple, psycho, pterodactyl,

pneumonia

• pronunciation depends on context

27

Decision Trees example

• Context: L1, L2, p, R1, R2

R1 = “h”

Yes No

P loopholeF physicsF telephoneF graphF photo

P peanutP payP appleø appleø psychoø psychoøpterodactyløpneumonia

28

Decision Trees example

• Context: L1, L2, p, R1, R2

R1 = “h”Yes No

P loopholeF physicsF telephoneF graphF photo

P peanutP payP appleø appleø psychoøpterodactyløpneumonia

Yes No

Ploophole

F physicsFtelephoneF graphF photo

L1 = “o”

R1 = consonantNoYes

PpeanutP pay

P appleø psychoø pterodactylø pneumonia

29

Decision Trees example

• Context: L1, L2, p, R1, R2

R1 = “h”Yes No

P loopholeF physicsF telephoneF graphF photo

P peanutP payP appleø appleø psychoøpterodactyløpneumonia

Yes No

Ploophole

F physicsFtelephoneF graphF photo

L1 = “o”

R1 = consonantNoYes

PpeanutP pay

P appleø psychoø pterodactylø pneumonia

try “PARIS”

30

Decision Trees example

• Context: L1, L2, p, R1, R2

R1 = “h”Yes No

P loopholeF physicsF telephoneF graphF photo

P peanutP payP appleø appleø psychoøpterodactyløpneumonia

Yes No

Ploophole

F physicsFtelephoneF graphF photo

L1 = “o”

R1 = consonantNoYes

PpeanutP pay

P appleø psychoø pterodactylø pneumonia

Now try “GOPHER”

31

Training a Decision Tree

• At each node, decide what the most useful split is.– Consider all features– Select the one that improves the performance

the most

• There are a few ways to calculate improved performance– Information Gain is typically used.– Accuracy is less common.

• Can require many evaluations

32

Pronunciation Models in TTS and ASR

• In ASR, we have phone hypotheses from the acoustic model, and need word hypotheses.

• In TTS, we have the desired word, but need a corresponding phone sequence to synthesize.

33

Next Class

• Language Modeling• Reading: J&M Chapter 4