Today's speech recognition systems require hundreds of hours of example data for training the acoustic models.  While such large corpora are available for the major languages, this is not the case for smaller languages, making them 'under-resourced'�. One of the underlying reasons for this data hungriness is that the dimensionality of feature vectors used in state-of-the-art speech recognition systems (typically in the range 30-40) is much larger than the intrinsic dimensionality of speech which is estimated to be 7-10 only. Efforts to make the intrinsic dimensionality smaller have been largely futile as the constraints are too complex for our by and large linear techniques. This inefficiency in basic representation, combined with other inefficiencies in mainstream context-dependent modeling makes that the hundreds of thousands parameters that constitute an acoustic model are largely redundant.      

The objective of this project is to apply novel mathematical techniques (e.g. spectral clustering) that can capture constraints - not in the feature space - but in the model space, i.e. in the underlying HMM parameters. Such constraints will lead to lesser requirements on the size of the training databases and should increase robustness in all situations where we don't have large corpora available, such as speaker adaptation, accent adaptation or modeling of under-resourced languages. Apart from general principles, two test cases will be  be studied in more detail : i) 'Afrikaans'�, for which data from Dutch and Flemish can be reused; ii) languages form the Bantu family as spoken in South Africa for which we can only bootstrap from a wide set of rather unrelated languages.

Qualifications

Candidates ideally have a university degree in engineering, computer science or applied mathematics. Skills and experience in any of the following areas are welcomed:

• speech recognition and speech modeling
• strong background in linear algebra and/or statistical parameter estimation
• some familiarity with Dutch or Afrikaans
• computational skills (MATLAB, C, UNIX, Python)

Position

Within this project there is funding for a 4yr Ph.D. scholarship.  Alternatively we will also accept applications for a 2 yr junior post-doc with significant relevant experience.

 Contact

Dirk Van Compernolle - compi@esat.kuleuven.be

Project Partners

This research will be carried out the K.U.Leuven, Belgium in the context of the AMODA project in collaboration with Council for Scientific and Industrial Research (CSIR), Pretoria, South Africa.