Ingen lysbildetittel

1
RUNDKASTAn Annotated NorwegianBroadcast News
Speech Corpus

• LREC 2008
• Ingunn Amdal, Ole Morten Strand,
• Jørn Almberg, and Torbjørn Svendsen

2
Overview

• Purpose of Rundkast
• An overview of the database Rundkast
• Structure of annotation
• Orthographic transcription
• Broad phonetic annotation

3
Purpose of Rundkast

• Databases of broadcast news can be used for a
  number of research topics in speech technology
  such as
• Supplement to existing databases of read speech
  for training and testing automatic speech
  recognition and speaker adaptation.
• Research on recognition of spontaneous speech.
• Research on automatic indexing of audio data.
• Research on topic and/or speaker segmentation.
• Research on speech/non-speech detection (e.g.
  background music).
• International research cooperation involving
  speech technology for broadcast news
  applications.
• A corpus of this kind is necessary for language
  technology research, but has not been available
  for Norwegian

4
Overview of Rundkasthttp//www.iet.ntnu.no/projec
ts/rundkast/

• Database of 77 hours radio broadcast news
  fromthe Norwegian Broadcasting Corporation
  (NRK)
• Read and spontaneous speech, as well as
  spontaneous dialogsand multipart discussions
• There is large variation between speakers,
  speaking styles and topics
• Speaker turns may be rapid and several speakers
  may talk simultaneously
• The quality of the recordings include studio and
  telephone(mobile, satellite etc)
• Frequent occurrences of background noise,
  jingles,music and audio illustrations
• Funded by the Norwegian University of Science and
  Technology (NTNU)

5
Structure of annotation

• Rundkast is hierarchically organizedand
  orthographically annotated
• Name of programme, type and date
• Name of speaker (if known) and dialect (5
  regions)
• Type of speech spontaneity, channel, recording
  quality
• Segmented in speaker turns of app. 2-5 seconds
• Orthographic transcription (standard Norwegian)
• Labels for noise (speaker noise, background noise
  etc.)
• Labels for pronunciation mistakes, foreign words,
  unintelligible speech etc.
• 70 hrs work per hour of recording
• Transcriber used for annotation standard-tool

6
Hierarchy of annotation levels
levels 1section, 2speaker turn, and 3segment
7
Orthographic transcription

• The lowest level in the annotation hierarchy,
  segments, are transcribed orthographically.
• Orthographic transcription of spoken language is
  a challenge, especially for Norwegian. Using
  dialect also in official circumstances is more
  and more accepted.
• The majority of RUNDKAST is not compliant to any
  standard pronunciation.
• The aim of the conventions for the orthographic
  transcription in RUNDKAST is to minimize
  uncertainty about pronunciations and facilitate
  consistency.

8
Orthographic transcriptionMain conventions

• Words are transcribed with the written forms
  closest to actual pronunciations. A limited
  number of interjections are allowed.
• Text codes are used to mark mispronunciations,
  truncations, and unknown words.
• Numbers and symbols are written out as words.
• Abbreviations are not used.
• Punctuation marks are restricted to comma,
  period, and question mark.
• Space is used between spelled letters, also when
  acronyms have spelled pronunciation.
• Capital letters are used in proper names,
  spellings, and acronyms, but not at the start of
  sentences.

9
Example annotation in Transcriber
10
Broad phonetic annotation

• Part of the data were to be phonetically
  annotated
• Use for low-level experiments in ASR (new
  methods), smaller Norwegian counterpart to TIMIT
• Auto-segmentation for e.g. unit selection TTS
• Annotation to be based on existing standards
  with necessary adjustments
• Exploit experience and specifications from
  development of Norwegian speech synthesis
  databases
• Suitable level of detail Acoustic boundaries
  should be labeled, but more phonemic than
  phonetic
• Consistency of utmost importance!

11
Broad phonetic annotationSelected data

• 10 speakers (5 male and 5 female)
• Amount of speech per speaker
• app 5 min planned speech and 1 min spontaneous
  speech
• discard noisy parts (as far as possible)
• from more than one programme
• use turn segmentation from orthographic
  annotation
• All in all 1 hour of speech
• Approximately 1000 hours of work

12
Broad phonetic annotationMain principles

• The annotation is mainly phonemic using the
  phoneme symbols closest to the perceived sound
• Acoustic boundaries should be marked some
  acoustically motivated symbols are included
• A transcription as close as possible to the
  citation form is preferred
• Norwegian standard SAMPA is preferred
• Some English phonemes included as well as dialect
  variants
• Example 3 variants of the /r/-sound/r/
  (tap/trill)/R/ (uvular fricative)/r\/
  (approximant)

13
Broad phonetic annotationAnnotation procedure

1. Conversion of orthographic transcription to a
   format suitable for automatic transcription.
2. Automatic segmentation with a phonotypical
   transcription using a speech recognizer.
3. Manual correction of both segments and labels by
   four phonetics students using Praat.
4. Format check.
5. Control of all annotation by one supervisor.

14
Broad phonetic annotationComments on deviations

• Always cases of uncertainty, need a log for
  these.
• Problem will the log be read?
• Solution Codes for deviations!
• Additional Praat tier for deviations
• Synchronous with the phoneme tier
• Easy to utilize automatically
• Examples
• creaky voice
• unexpected voiced/unvoiced
• uncertain boundary or symbol
• ... in addition a log file with whatever
  deviations left

15
Example annotation in Praat
16
Concluding remarks

• Availability
• Planned to be included for non-commercial use in
  a future Norwegian language bank
• Will complement other corpora also intended to be
  included
• To be validated by Spex
• Planned use at NTNU SIRKUS project
• Investigation in new paradigms for ASR
• Low-level phone recognition experiments initially
• multi-linguality aspects
• Spoken information retrieval