The 10-milion-words Spoken Dutch Corpus and its potential use in experimental phonetics

1
The 10-milion-words Spoken Dutch Corpus and its
potential use in experimental phonetics

• Louis C.W. Pols
• Institute of Phonetic Sciences
• University of Amsterdam

100 Years of Experimental Phonetics in
Russia St.-Petersburg State Univ., Febr. 1-4, 2001
2
Amsterdam city center
Herengracht 338
3
Overview

• Introduction
• Corpus design, recording, digitization
• Orthographic transcription
• Part-of-speech tagging, lemmatization and
  syntactic annotation
• Phonetic transcription
• Prosodic transcription
• Exploration
• Potential phonetic benefit

4
Introduction

• appropriate topic given long Russian tradition
• Dutch-Flemish initiative
• 10 Mƒ, 10 M words (about 1000 hrs of speech)
• start June 1998, 5 yrs, 7 releases (audio ann.)
• many speaking styles, also over telephone, only
  adult speakers, ABN variants but no dialect
• for linguistics and speech/language technology
• rights with NTU (http//www.taalunie.nl)

5
Corpus design(number of words x 1000)
dialogues and multilogues
monologues
6
Recording, digitization

• mono or stereo using portable DAT-recorders
• 16 kHz and 16 bit (telephone recordings at 8
  kHz and 8 bit)
• .WAV format in PRAAT
• meta data about recording and speaker
• 7 audio releases on CD-ROM, or DVD (future?)
• annotations updated with each release

7
Orthographic transcription (1)

• by trained students, checked by expert
• according to fixed protocol no text
  interpretations
• transcr. aligned at few sec. chunks multiple
  tiers
• few punctuations capitals for names only
• standard spelling conventions, checked vs.
  lexicon
• special mark-up symbols
• d dialect words z regionally accented words
• t interjection a truncated wrd u
  mispronunciation
• v foreign words n new words x hardly
  intelligible
• ggg speaker sounds xxx unintelligible
  word(part)(s)

8
Orthographic transcription (2)
9
Part-of-speech tagging

• all words in the text automatically tagged
• discontinuous verbs not recognized at this level
• Dutch tag set with 10 major word classes
• (noun, adjective, verb, pronoun, article,
  numeral, preposition, adverb, conjunction, and
  interjection)
• additional morpho-syntactic features per class
• (e.g., singular, dimunitive and neuter for nouns)
• resulting in some 300 tags
• self-learning automatic tagger (given context)

10
Lemmatization

• all words autom. paired with base form (lemma)
• verbs ? infinitive (gedaan ? doen) other forms
  ? stem (vijfde ? vijf) truncated forms ? full
  forms (zn ? zijn)
• base form must be an independently existing form
• (hersenen ? hersen meisje ? meis)
• discontinuous verbs and split prepositions are
  not recognized at this level (op...bellen
  van...uit)
• one and only one baseform per word
• (vliegen ? verb vliegen, or noun vlieg, depending
  POS)

11
Broad phonetic transcription (1)

• on 10 of the data (mainly dialogues)
• hand correction of automatic phonetic
  transcription
• across-word assimilation, levels of reduction?
• use of extended SAMPA
• within PRAAT
• word level respected
• die ik wel vind dat ze kloppen ? di k wEl
  fInt_tAt s_at_ klOp_at_
• no hand segmentation at phoneme level

12
Broad phonetic transcription (2)
13
Signal coupling, word alignment

• the phonetically transcribed part (1 M words)
  will be automatically aligned at word level
• using ASR techniques (forced alignment)
• this word alignment will be hand corrected
• pauses and noises will also be aligned
• geminate plosives are aligned separately, others
  shared (komt terug ? kom t erug is zeker ?
  isseker)
• inserted phonemes are shared with neighbouring
  words (toen belde n ie naar huis ? belden nie
• all the rest may be automatically aligned only
• few seconds chunks are always accessible

14
Syntactic annotation

• 10 will be semi-automatically annotated
• procedure still under developed
• interactive annotation software from NEGRA
  project (Saarbrücken) will be used
• taking into account idiosyncracies of speech,
  such as hesitations, false starts, clause
  extensions
• functional information (dependency labels)
• category information (in form of node labels)

15
Prosodic annotation

• manually, on 250K words subset only
• procedure still under development
• prosodic markers in orthography
• 1) prosodic boundaries
• long silences (?)
• phrase boundaries (?)
• other discontinuities, like (filled) pauses ()
• 2) prominence ( before vowel in prominent
  syllable)
• sp. A nee ? Jan heeft negen medailles ?
  zeven medailles. ?
• sp. B zeven ?

16
Exploration software

• COREX tool under developed (Max Planck Inst.)
• both locally and internet-based (Java)
• 1) browser
• 2) viewer for orthography and annotations, plus
  waveform display and audio player (time synchr.)
• 3) search module, also on meta data

17
Potential phonetic benefit

• huge database, many speakers/styles,real speech
• easily accessible via orthography, plus audio
• partly accessible via phonetic transcription
• no segmentation at phoneme level (automatic?)
• automatic segmentation at word level
• after COREX search own additions possible
• f.i. spectro-temporal analyses via PRAAT scripts
• f.i. svarabhakti vowel, final n-deletion,
  assimilation
• f.i. vowel reduction, turn-taking behavior, etc.

18
More information

• see references in paper
• see websites mentioned in paper
• second release Oct. 2000
• new releases every half year
• feedback from users group (workshops)
• useful for proposed INTAS project Spontaneous
  speech of typologically unrelated languages
  (Russian, Finnish and Dutch) Comparison of
  phonetic properties (De Silva, 2000)