Centre for Computational Linguistics: Projects

ATraNoS: Automatic Transcription and Normalisation of Speech

Time Span: 2000 - 2004
Institute for Innovation in Science and Technology (IWT), STWW program
F. Van Eynde, V. Vandeghinste, P.Dirix
Other participants: ESAT/PSI (K.U.Leuven), ELIS (R.U.Gent), CNTS (U.Antwerpen)

The CCL has two tasks in this project. The first task concerns the reduction of the number of Out-of-Vocabulary words by building a large scale lexicon for the speech recognizer that takes the existence of compound words in Dutch into account. The second task concerns the 'translation' of the transcribed speech input into 'subtitle Dutch'.

1. Reduction of Out-of-Vocabulary words.

One of the main reasons for speech recognition errors is the occurence of Out-of-Vocabulary words. In Dutch, word compounding is a productive process and hence it is impossible to give a list of all Dutch words that should be in the recognizers lexicon.
A lexicon was built that consists of a list of non compound words, together with a list of compound parts (called the Quasi-word List) and a perl module which allows for online word compounding. The perl module makes use of a ruleset describing the compounding rules for Dutch. As the rules tend to overgenerate, a statistic measure estimating the likeliness of the compound is introduced to get a better accuracy in compounding.

2. Translation of transcribed speech into Subtitle Dutch

Subtitles are not a transcription of what is said in a TV-program. Often some parts of what is being said are left out. This can be due to the fact that the speaking rate is too fast in order to put all the words on the screen and still comply to the 6-seconds rule. So, a system is designed that parses the incoming sentence and decides what parts of the input sentence should be retained in the output sentence.

CCL

Layout: webmaster@ccl.kuleuven.ac.be
Information Provider: Centrum voor Computerlinguïstiek
Comments to the Webmaster: Ineke.Schuurman@ccl.kuleuven.ac.be