Phonemes and words segmentation

Definition

• the process of taking the orthographic transcription text of an audio speech segment, like IPUs, and determining where particular phonemes/words occur in this speech segment.
• Definition:
  • IPUs = Inter-Pausal Units

Data preparation: example

Expected result

• Time-aligned phonemes and tokens, including events like noises or laughter

Phonemes and words segmentation: research approach of SPPAS

1. text normalization
2. phonetization (grapheme to phoneme conversion)
3. alignment (speech segmentation)

All three tasks are fully-automatic, but each annotation result can be manually checked.

Text Normalization

• Any system that deals with unrestricted text needs the text to be normalized.
• It’s the process of segmenting a text into tokens.
• Automatic text normalization is mostly dedicated to written texts, in the NLP community
  • Text normalization development is commonly carried out specifically for each language and/or task even if this work is laborious and time-consuming. Actually, for many languages there has not been any concerted effort directed towards text normalization.

Text Normalization: SPPAS approach

• A generic approach:
  • a method as language-and-task independent as possible.
  • this enables adding new languages quickly when compared to the development of such tools from scratch.
• This method is implemented as a set of modules that are applied sequentially to the text corpora.
• The portability to a new language consists of:
  • inheriting all language independent modules;
  • (rapid) adaptation of other language dependent modules.

Text Normalization main steps

1. Split:
   • use whitespace or characters to split the utterance into separated strings
2. Replace symbols by their written form:
   • based on a lexicon
     • ° is replaced by degrees (English), degrés (French), grados (Spanish), gradi (Italian), mức độ (Vietnamese), 度 (Chinese), du (Chinese pinyin and Taiwanese)
     • ² is replaced by square (English), carré (French), quadrados (Spanish), quadrato (Italian), bình phương (Vietnamese), 平方 (Chinese), ping fang (Chinese pinyin)

Text Normalization main steps (continued)

1. Segment into words:
   • fixes a set of rules to segment strings including punctuation marks
   • based on a lexicon and rules
     • aujourd’hui, c’est-à-dire
     • porte-monnaie, cet homme-là, voulez-vous
     • poudre d’escampette, trompe-l’oeil, rock’n roll

Text Normalization of speech transcription

• Speech transcription includes speech phenomena like:
  • specific pronunciations: [example,eczap]
  • elisions: examp(le)
• Then two types of transcriptions can be automatically derived by the automatic tokenizer:
  1. the “standard transcription” (a list of orthographic tokens/words);
  2. the “faked transcription” that is a specific transcription from which the obtained phonetic tokens are used by the phonetization system.

Text Normalization of speech transcription: example

• Transcription:
  • mais attendez et je mes fixations s(ont) pas bien réglées [c’,z] est en fait c’est m- + ma chaussure qu(i) était partie
• Tokens standards:
  • mais attendez et je mes fixations sont pas bien réglées c’ est en_fait c’est ma chaussure qui était partie
• Tokens:
  • mais attendez et je mes fixations s pas bien réglées z est en_fait c’est m- + ma chaussure qu était partie

Text Normalization: current languages

• French: 347k words
• Italian: 389k words
• German: 383k words
• Polish: 576k words
• English: 121k words
• Catalan: 93k words
• Portuguese: 41k words
• Spanish: 22k words
• Japanese: 18k words
• Mandarin Chinese: 110k words
• Cantonese: 46k words
• Korean: 33k words
• Min nan: 1k syllables in pinyin
• Naija: 5k words + English

The better lexicon, the better automatic text normalization.

Text Normalization: Adding a new language

1. add lexicons
2. add the num2letter module

Example:

 Roxana Fung, Brigitte Bigi (2015).
 Automatic word segmentation for spoken Cantonese.
 In Oriental COCOSDA and Conference on Asian Spoken Language Research and Evaluation,
 pp. 196–201.
 
 

Text Normalization: reference

 Brigitte Bigi (2014).
 A Multilingual Text Normalization Approach.
 Human Language Technologies Challenges for Computer Science and Linguistics.
 LNAI 8387, Springer, Heidelberg. ISBN: 978-3-319-14120-6. Pages 515-526.
 

Phonetization

• Phonetization is also known as grapheme-phoneme conversion
• Phonetization is the process of representing sounds with phonetic signs.
• Phonetic transcription of texts is an indispensable component of text-to-speech (TTS) systems and is used in acoustic modeling for automatic speech recognition (ASR) and other natural language processing applications.

Converting from written texts into actual sounds, for any language, cause several problems that have their origins in the relative lack of correspondence between the spelling of the lexical items and their sound contents.

Phonetization: SPPAS research approach

• a generic approach:
  • consists in storing a maximum of phonological knowledge in a lexicon.
  • In this sense, this approach is language-independent.
• the phonetization process is the equivalent of a sequence of dictionary look-ups.

Phonetization: dictionary

• An important step is to build the pronunciation dictionary, where each word in the vocabulary is expanded into its constituent phones, including pronunciation variants.

Phonetization of normalized speech transcription

• SPPAS is proposing a language-independent algorithm to phonetize unknown words
  • given enough examples (in the dictionary) it should be possible to predict the pronunciation of unseen words purely by analogy.
• Example with the unknown word “pac-aix”:
  • English: p-{-k-aI-k-s|p-{-k-eI-aI-k-s|p-{-k-aI-E-k-s|p-{-k-eI-aI-E-k-s
  • French: p-a-k-E-k-s
  • Mandarin Chinese: p_h-a-a-i

Phonetization: example

• Tokens:
  • mais attendez je
• Phonetization:
  • m-E-z|m-e|m-E|m-e-z|m A/-t-a~-d-e|A/-t-a~-d-e-z e|E Z|Z-eu|S

Phonetization: current languages

• French: 652k entries
• English: 121k entries
• Italian: 590k entries
• German: 438k entries
• Spanish: 24k entries
• Catalan: 94k entries
• Portuguese: 43k entries
• Polish: 300k entries
• Japanese: 20k entries
• Mandarin Chinese: 114k entries
• Cantonese: 59k entries
• Korean: 128 entries (!) is under construction
• Min nan: 1k entries
• Naija: 4k entries

The better dictionary, the better automatic phonetization.

Phonetization: reference

 Brigitte Bigi (2016).
 A phonetization approach for the forced-alignment task in SPPAS.
 Human Language Technologies Challenges for Computer Science and Linguistics.
 LNAI 9561, Springer, Heidelberg. 
 

Alignment

• Alignment is also called phonetic segmentation
• The alignment problem consists in a time-matching between a given speech unit along with a phonetic representation of the unit.
• Many freely available tool boxes, i.e., Speech Recognition Engines that can perform Speech Segmentation
  • HTK - Hidden Markov Model Toolkit
  • CMU Sphinx
  • Open Source Large Vocabulary CSR Engine Julius
  • …

Alignment

• Algorithms are language independent
• An acoustic model must be created
  • training procedure
  • based on examples
• “more data is good data”… or not!
• Training a model:
  • requires audio files
  • requires orthographic transcription
  • requires IPUs/utterrances segmentation

Alignment: current languages

• French
• English (from voxforge.org)
• Italian
• Spanish
• German
• Catalan
• Polish
• Mandarin Chinese
• Min nan
• Naija
• Japanese (from Julius software)
• Cantonese (from University of Hong Kong)
• Korean (under construction)

Alignment: results of French

• Unit Boundary Position Accuracy (UBPA): what percentage of the automatic-alignment boundaries are within a given time threshold of the manually aligned boundaries.

Alignment: references

 Brigitte Bigi (2012).
 The SPPAS participation to the Forced-Alignment task of Evalita 2011.
 B. Magnini et al. (Eds.): EVALITA 2012, LNAI 7689, pp. 312-321. Springer, Heidelberg.

 Brigitte Bigi (2014).
 The SPPAS participation to Evalita 2014.
 In Proceedings of the First Italian Conference on Computational Linguistics CLiC-it 2014
 and the Fourth International Workshop EVALITA 2014, Pisa, Italy.

 Brigitte Bigi (2014).
 Automatic Speech Segmentation of French: Corpus Adaptation.
 In 2nd Asian Pacific Corpus Linguistics Conference, pp. 32, Hong Kong.
 

Speech segmentation: main reference

 Brigitte Bigi, Christine Meunier (2018).
 Automatic speech segmentation of spontaneous speech.
 Revista de Estudos da Linguagem.
 International Thematic Issue: Speech Segmentation.
 Editors: Tommaso Raso, Heliana Mello, Plinio Barbosa,
 e - ISSN 2237-2083
 

• Click here to access the paper: http://www.periodicos.letras.ufmg.br/index.php/relin/article/view/13026

You reached the end of this tutorial!

Click here to go back to SPPAS tutorials web page

Syllables segmentation

Data preparation

• Requires time-aligned phonemes

Expected result

• Time-aligned syllables

• IMPORTANT: Syllabification was fully re-implemented in version 1.9.7 and the displayed result has not been updated

Syllabification: SPPAS approach

• A rule-based system to cluster phonemes into syllables
• Rules available for:
  • French
  • Italian
  • Polish
• This phoneme-to-syllable segmentation system is based on two main principles:
  1. a syllable contains a vowel, and only one;
  2. a pause is a syllable boundary.

Syllabification: SPPAS approach (continued)

• Phonemes are grouped into classes
• Classes for both French and Italian:
  • V - Vowels,
  • G - Glides,
  • L - Liquids,
  • P - Plosives,
  • F - Fricatives,
  • N - Nasals.

Syllabification: SPPAS approach (continued)

• Fix rules to find the boundaries between two vowels
  • general rules with V and C
    • VCV => V.CV
    • VCCV => VC.CV
  • exception rules based on the classes
    • VLGV => V.LGV
    • VPGV => V.PGV
  • other rules based on the phonemes themselves

Syllabification: resources

• A file with:
  • the list of phonemes and the corresponding class
  • the list of all rules
• Current languages:
  • French: 10 exception rules; 6 phoneme-based rules
  • Italian: 11 exception rules; 1 phoneme-based rule
  • Polish: 50 exception rules; 337 phoneme-based rules

Syllabification of Italian: reference

 Brigitte Bigi and Caterina Petrone (2014).
 A generic tool for the automatic syllabification of Italian.
 In Proceedings of the First Italian Conference on Computational Linguistics CLiC-it 2014 and
 of the Fourth International Workshop EVALITA 2014, pp. 73–77, Pisa, Italy.
 

Syllabification of Polish: reference

 Brigitte Bigi and Katarzyna Klessa (2015).
 Automatic Syllabification of Polish.
 In 7th Language and Technology Conference: Human Language Technologies as a Challenge for
 Computer Science and Linguistics, pp. 262–266, Poznan, Poland.
 

You reached the end of this tutorial!

Click here to go back to SPPAS tutorials web page

Repetitions detection

Repetitions

• Other-repetition is a device involving the reproduction by a speaker of what another speaker has just said.
• Other-repetition has been identified as an important mechanism in face-to-face conversation through their discursive or communicative functions.
• Example:

 AB: ils voulaient qu'on fasse un feu d'artifice en fait dans un voy- un foyer un foyer catho
 un foyer de bonnes soeurs
 CM: un feu d'artifice
 AB: ah ouais
 CM: dans un foyer de bonnes soeurs
 CM: @
 

Repetitions

• No system already existing
• Our solution is based only on lexical criteria
  • observable cues: time-aligned tokens converted to lemmas (if any).
• The system was used to propose a lexical characterization of OR: various statistics was estimated on the detected OR.

Other-Repetitions: result

Repetitions: reference

 Brigitte Bigi, Roxane Bertrand, Mathilde Guardiola (2014).
 Automatic Detection of Other-Repetition Occurrences: Application to French Conversational Speech.
 In Proceedings of the Ninth International Conference on Language Resources and Evaluation,
 pp. 836-842, Reykjavik, Iceland.
 

You reached the end of this tutorial!

Click here to go back to SPPAS tutorials web page