Long span units in example based speech recognition

50000403v1816
Promoter: Dirk Van Compernolle

details
Description: The ESAT speech group has been the leading institute in reviving interest in example based speech recognition. Over the past years we have developed a state-of-the-art hybrid speech recognition system, combining Hidden Markov Models (HMMs) and Example based concepts. While HMMs are a very powerful mathematical concept, they may not be 100% appropriate for modeling speech.
The example based methods on the other hand are better in modeling transients and in integrating certain long span consistencies; however, they may be weaker in terms of generalization.

In this project we want to tackle truly long span units and fuse acoustic-phonetic knowledge at the phone, syllable and word level.
Together with dealing with these longer span units at the recognition level,
we also intend to explore different signal representations which are better suited for these long time spans.

We are currently looking for a PhD student to strengthen the team and to deeply engage in one of the aspects of this research. This project will be carried out in close collaboration with other ongoing projects, such as the FWO project TELEX and the Marie Curie network Sound to Sense.

We are looking for a motivated candidate with both theoretical and experimental skills. Ideally he/she has a Master's degree in engineering or computer science. Previous experience in speech recognition is not necessary, though knowledge of or experience in one the following areas form an asset:
- speech recognition and speech modelling
- pattern matching
- programming experience in C or Python


Key words: speech recognition, statistical pattern matching,

Latest application date: 2010-04-25

Financing: available

Type of Position: scholarship

Duration of the Project : 4 years

Link: http://www.esat.kuleuven.be/psi/spra...elex_jobad.php

Research group: Department of Electrical Engineering (ESAT)

Remarks:


Apply to Click here to apply to this project