Aucun titre de diapositive

1
Quick Rich Transcriptions of Arabic Broadcast
News Speech Data
Niklas Paulsson, Djamel Mostefa, Chomicha
Bendahman ELDA (Evaluations and Language
resources Distribution Agency) Meghan Glenn,
Stephanie Strassel LDC (Linguistic Data
Consortium)
2
Outline

• Overview
• Transcription method
• Sources
• Collection
• Selection
• Transcription
• Segmentation, Sentence Units, Overlapping Speech,
  Markup
• Quality Control
• Conclusion

3
Overview

• Broadcast News Transcripts
• Arabic (MSA, MCA)
• Sources radio TV, mostly Middle East
• Verbatim orthographic transcripts, time-aligned,
  minimal mark-up
• QRTR to reduce time

4
Transcription Method (1)

• QRTR Quick Rich Transcription
• (QTR / QRTR / CTR)
• Amount of detail in markup
• Number of features identified
• Degree of accuracy
• Completeness
• Amount of time
• Number of quality checks

5
Transcription Method (2)
Type QTR QRTR CTR
Markup 3 5 7
Speaker noise 2 4 8
External noise 2 1 6
Sections - 3 3
Themes - - 10
Speaker turns - 3 4
Speaker info - 3 6
SU Type - 4 0
features 7 23 47
6
Sources (1)

• Two types of recordings
• Broadcast news (BN) talking head style news
  reports
• Broadcast conversation (BC) more interactive,
  talk shows, interviews, call-in programs,
  roundtable discussions
• Mainly MSA from Middle East
• MCA from North Africa and Middle East
• Overlapping speech
• 30 60 minutes of recordings
• collected from TV and Radio sources

7
Sources (2)
8
Collection

• Sources recorded from satellite
• Daily and weekly recordings
• Records video stream
• Audio extracted from video
• Saved in WAV or SPH
• 16 bits, 16 kHz

9
Audit

• Manual audit of all programs
• Procedure
• Listen to 30 sec samples of 3 sections
  beginning, middle, end
• Auditors can listen to additional segments if
  necessary
• Fills in a form for auditing the recordings
• Web-based auditing interface
• Checks
• Is there a recording?
• Is the audio quality ok?
• What is the language?
• Is it speech from the right program?
• What is the data type?
• What is the topic?

10
Selection

• Recordings rejected
• poor quality
• wrong language
• Passed audit eligible for transcription
• Criteria based on
• data amount
• sources
• dates
• 2000 hours in 24 sets
• Sent in 20 300 hours packages for transcription
• Period Apr. 2004 Aug. 2007

11
Transcription

• Orthographic, verbatim transcripts
• Arabic script
• No vowels
• Segmented and time stamped
• Speaker names
• Sentence Units
• Noise markers
• Overlapping speech
• Foreign language markup

12
XTrans (1)

• Tool for broadcast news and conversation
• Multi-lingual (UTF-8)
• Multi-platform (Windows, Linux, FreeBSD)
• Output TDF format
• Compatible with Transcriber format

13
XTrans (2)
14
Segmentation (1)

• Segment data into sections
• Speech delimited by pause or silence
• Non speech sections music, silence, ads, etc
• Lasts 5 20 seconds
• Sections are classified as
• News report (BN)
• Conversation (BC)
• Non-news
• Sections are next grouped into speaker turns
• Single speaker or overlapping turn
• Statement Units (SU)
• Speaker ID or name for each turn

15
Sentence Units

• Group utterances into clusters of words
• Each cluster represent a sentence-like unit
• Each unit receives a label
• Statement
• Question
• Incomplete
• Non-Speech

16
Overlapping Speech

• Many recordings include conversations
• Portions of speech that are overlapping
• Segmented and annotated
• No SU type
• Could be quite challenging
• Difficult portions annotated as non-speech

17
Markup (1)

• Minimal set of markers
• Hesitations
• Truncated words
• Mispronunciations
• Made up words
• Noise
• Difficult speech

18
Markup (2)

• Noise markers
• Background noise
• Speaker noise laugh, cough, sneeze, lipsmack
• Dialect / language markup
• Non-MSA (MCA)
• English
• French
• Foreign Language

19
Quality Control

• Limited quality control due to time constraints
• Quick Verification procedure
• Max 18 min / file
• Focus
• Transcription matches speech
• Segmentation
• Speaker names
• Orthography
• Procedure
• Checks 3 segments of 3 min beginning, middle and
  end
• Transcriptions that did not pass sent back to
  transcribers

20
Conclusion

• Arabic broadcast data
• gt2000 hours transcribed
• 330k words
• Useful for quantative manual transcripts
• Limited timeframe
• Minimal but useful markup
• Quality control
• Training ASR systems for MT

21

• Thanks for your attention