Overview

Point of departure

We have a new corpus of casual spoken Czech with manual phonetic transcription → ORTOFON v1

Goals

1. Here’s some cool stuff you can do with it.
2. And by you, I mean you! 👉 😱 👈

Resources

Google: korpus ortofon

This presentation: https://trnka.korpus.cz/~lukes/slides/slavicorp2018/ortofon

Our wiki: https://wiki.korpus.cz/doku.php/cnk:ortofon

KonText query interface: https://kontext.korpus.cz/first_form?corpname=ortofon_v1

All the data via LINDAT:

• ELAN transcripts + audio: http://hdl.handle.net/11234/1-2579
• vertical: http://hdl.handle.net/11234/1-2580

Corpora of spoken Czech at the CNC I

Corpora of spoken Czech at the CNC II

Geographically

Some figures

Per-document metadata

• 12 pre-defined situation types
• year, month and location of recording
• relationship between speakers in recording
• gender mix of speakers in recording
• generation mix of speakers in recording
• … and more.

Per-speaker metadata: balancing

On the basis of the following metadata:

• gender: male × female
• age: under 35 × over 35
• education: tertiary × other
• childhood dialect region of residence: 10 regions

Resulting number of categories: \(2 \times 2 \times 2 \times 10 = 80\)

Ideally: equal representation of these 80 categories, at least 5 speakers per category.

→ Target number of words per category: \(\frac{1\ 000\ 000}{80} = 12\ 500\)

Transcribed using ELAN

Accessible via KonText (but also LINDAT)

Phonetic transcription

• simplified phonetic transcription
  • use regular Czech alphabet graphemes as much as possible
  • no phonetic diacritics
• stress group boundaries
• aligned with basic transcript: tokens match \(1:1\)

Assimilation of voicing in Czech

• final devoicing:

  <hrad (nom.), hradu (gen.)> → [hrat, hradu]

• regressive / anticipatory assimilation of voicing, even across word boundaries:

  <hrad, hrad byl> → [hrat, hrad bil]

• triggered by voiced obstruents

• in Moravia/Silesia (Eastern part of the country), also triggered by sonorants [r, l, m, n, j…]:

  <tak jako> → [tag jako]

Traditional dialect regions

Assimilation of voicing ~ Region of childhood residence

Which forms assimilate to sonorants?

# of variants ~ Frequency (\(\rho = 0.76\))

In KonText: protože

In KonText: samozřejmě

In KonText: normálně

Identify competition with entropy

Formally reduced pronunciations

• motivated by:
  • frequency, length (across languages)
  • lexical effects
• edit distance between abc and zbc:
  • Levenshtein: 1
  • normalized Levenshtein: 0.33

• in practice, deletion and substitution (~ formal simplification) much more common than addition (epenthesis)

  → high normalized Levenshtein distance ~ high amount of simplification

Normalized Lev. dist. ~ Frequency

Instead of a conclusion…

Google: korpus ortofon

This presentation: https://trnka.korpus.cz/~lukes/slides/slavicorp2018/ortofon

Our wiki: https://wiki.korpus.cz/doku.php/cnk:ortofon

KonText query interface: https://kontext.korpus.cz/first_form?corpname=ortofon_v1

All the data via LINDAT:

• ELAN transcripts + audio: http://hdl.handle.net/11234/1-2579
• vertical: http://hdl.handle.net/11234/1-2580

Acknowledgments

This research was supported by the Czech National Corpus project (LM2015044) funded by the Ministry of Education, Youth and Sports of the Czech Republic within the framework of Large Research, Development and Innovation Infrastructures.

Slides: https://trnka.korpus.cz/~lukes/slides/slavicorp2018/ortofon