Collecting and Transcribing Real Chinese Spontaneous Telephone Speech Corpus

1
Collecting and Transcribing Real Chinese
Spontaneous Telephone Speech Corpus

• Limin Du, Chair Professor
• Director, Center for Speech Interactive
  Information Technology
• Institute of Acoustics, Chinese Academy of
  Sciences
• October 21, 2000

2
Background

• Spontaneous speech interactive via telephone is a
  very prospect application, building speech
  recognition systems in terms of the variations in
  acoustics and spoken styles for telephone
  application is necessary
• There is no large-scale Chinese Spontaneous
  Telephone Speech Corpus available for research
• Simulating telephone speech corpus (1997, C-SIIT,
  IOA, CAS)
• Microphone speech corpus pipeline to telephone
  telephone speech
• Collecting real telephone speech data seems to be
  a formidable task
• Laws
• Costs
• Chinese-English speech translation (CEST)
  project, an collaboration between CAS-ATT
  (1998-2003) is an strong driving for this work

3
Real Telephone Speech Collection

• A dialogue oriented collection paradigm
• Human-Human conversations
• Human-machine dialogues

4
Speech Data Processing

• Sampling
• 8kHz sampling
• 16bits A/D quantization
• Utterance Segmentation
• One Speaker switching for one utterance
• Utterances in average length of 3 seconds

5
Speech Data Transcribing

• What to Label?
• How to Label?

6
What to Label?

• Information about Speakers and Environments
• speakers dialect, mood, gender, speech quality
• Transcribing
• Chinese characters
• Pinyins
• Other acoustic event labels
• laugh, lip smack, throat clearing, breath, cough,
  filled pauses, telephone adjusting, background
  speech, etc.
• Time Stamp
• Other acoustic event are bracketed with time
  stamps automatically when transcribing with a
  special software tool

7
Detailed Issues Concerned

• Mispronunciation
• Mispronunciation often occurs in daily life. For
  example the speaker probably read Chinese
  character ? (whos correct pronunciation is
  shan1) as san2. In such a case, the
  associated speech segment is transcribed as
  ?(san2) to present the right text and real
  pronunciation
• Numbers
• Arabia representation of numbers is a natural
  method, but it cannot be mapped to a single
  pronunciation. So, transcribers are required to
  transcribe all numbers with Chinese characters

8
Other Acoustic Events

• ?? ???? ????
• PAUSE1 AI UH
• PAUSE14 AI UH
• PAUSE12 A UNG
• PAUSE33 KA A UNG
• PAUSE20 ANG UNG
• PAUSE26 ANG UNG
• PAUSE19 AN EN
• PUASE4 CHA AO
• PAUSE18 GAN UH
• PAUSE21 HE EN
• PAUSE27 NE EN
• PAUSE22 YUN UM
• PAUSE34 LENG UH
• PAUSE15 TONG UH

9
Other Acoustic Events(cnt)

• ?? ???? ????
• PAUSE31 NONG EN
• PAUSE17 HEN EN
• PAUSE24 EN EN
• AA
• AI
• EN
• UH
• AO
• SIL ???
• NOISE
• LAUGH
• ANG BREATH ??
• HESITATION ??

10
Transcription Example

• ltBeginStamp 0gtFILLERltEndStamp 257gt ltBeginStamp
  260gt NOISE ltEndStamp 928gt???????????????????
  ??????ltBeginStamp 5933gt FILLERltEndStamp
  6250gtltBeginStamp 6228gt FILLERltEndStamp 6386gt??

11
How to Label?

• Improving transcribers efficiency reducing the
  possibility to generate errors
• A labeling tool developed specially for this
  task.
• Training transcribers
• Usually our employees assisted speech research
  for more than one year and with good working
  records
• Part time employees trained by our employees
  before working at

12
Statistical Results in GeneralChinese
Spontaneous Telephone Speech Corpus (CSTSC)

13
Statistical Results in Details 180 human-human
dialogues, 38 human-machine dialogues

14
Summary

• C-SIIT, CAS started the work to build telephone
  speech corpora under very limited budget 3 years
  ago
• The efforts and experiences in collecting real
  Chinese telephone speech corpus are introduced
• C-SIIT will continue the Activity on Real Chinese
  Telephone and Mobile phone Speech Corpora and try
  best to make most of the corpora already built
  ,in building, in planning, released to public
• Suggestions and commences from all of you are
  appreciated
• Thanks!