Possible databases for use at WS06

This is a listing of the databases that have been mentioned for possible use at WS06.

Audio only

• SVitchboard (currently, this is the main database we plan to use)
  • Small-vocabulary (10- to 500-word), conversational speech extracted from the Switchboard database
  • Described in [King et al., Interspeech 2005]
  • Also see the SVitchboard project site

• PhoneBook (a possible alternative in case we want to look at isolated words, or just a somewhat easier task)
  • ~8000-word vocabulary, isolated words, phonetically rich
  • Described in [Pitrelli et al., ICASSP 1995]
  • Also see the PhoneBook database page

Audio-visual

• AVTIMIT (MIT)
  • Multi-speaker
  • Read speech consisting of sentences from the TIMIT database
  • Described in Hazen et al., ICMI 2004

• AVICAR (UIUC)
  • Multi-speaker
  • Read speech consisting of isolated digits and letters, phone numbers, and sentences
  • In-car environment with varying car noise levels
  • Described in Lee et al., ICSLP 2004
  • See the AVICAR project web site
  • Manual face and lip segmentations: attached at bottom of page
  • Description of the AVICAR Data on ws06afsr

• CMU audio-visual speech processing project data (CMU)
  • 10 speakers
  • Isolated words
  • Lip tracking parameters available
  • See the project web site

• CUAVE (Clemson)
  • 36 speakers, each reading 60 isolated digits and 60 connected digits
  • Also a number of speaker-pair recordings
  • Studio environment
  • See RelevantPapers for some example results

• New continuous numbers database (MIT)?
  • Stay tuned...

Transcribed data

For sanity checks, or in order to do separate acoustic/pronunciation modeling experiments

• STP (Switchboard Transcription Project)
  • Manually transcribed (for the most part) utterances from the Switchboard database
  • Transcribed at a detailed phonetic level (including diacritics, e.g. nasalization, frication)

• New data transcribed for the workshop

Articulatory measurement data

For detailed study of articulatory phenomena. The current thinking is that we will not use these, as they lack some useful articulatory information and are perhaps too difficult to translate to the discrete feature values that we intend to use. However, this is up for discussion.

• University of Wisconsin X-ray microbeam database, a multi-speaker database described in this user's handbook
• MOCHA, consisting of electromagnetic articulograph (EMA) recordings of 2 speakers

Discussion area for this page:

-- KarenLivescu - 08 Dec 2005

Attachment	Action	Size	Date	Who	Comment
avicar_faces_and_lips.tgz	manage	6127.3 K	04 Jun 2006 - 13:24	MarkHasegawaJohnson