4 March: fyi -- phd funding, Edinburgh
Index of March 2005 | Index of year: 2005 | Full index
speech recognition (ASR). Although they are seemingly very simple models,
in fact they require a complex system of context-dependent models,
parameter sharing and adaptation algorithms to achieve the best performance.
HMMs are a member of a wider family of models - dynamic Bayesian networks
(DBNs). There are an infinite variety of other DBNs waiting to be tried for
ASR. DBNs can be formulated to reflect our understanding of the speech
signal; one example of this would be multi-streamed DBNs (such as the
factorial HMM) in which the factors have explicit linguistic
interpretations - the factors might represent aspects of the speech
production process.
Dependencies can be introduced between these hidden factors, creating ever
richer model structures (at the cost of increased computational
complexity). The goal is to find model structures that improve recognition
accuracy whilst remaining computationally feasible.
We have already started exploring various forms of DBN, but there is still
a lot of scope for exciting and original research in this area. The
toolkits and computer power are now available to work with models that were
intractable until recently. It may be that some of the techniques developed
for HMMs can be transferred to DBNs, or we could build things like
parameter tying, pronunciation variation, language modeling and adaptation
into the model structure itself.
Full funding (fees plus living allowance) for eligible UK or EU students is
available.
Address for Applications:
Simon King
Simon.King@ed.ac.uk
-----------------------------------------------------------
LINGUIST List: Vol-16-640
----- End forwarded message -----
Index of March 2005 | Index of year: 2005 | Full index