NEW BOOK: Applications of NNs to speech understanding

[Ralf Kompe]

(Sorry, if you receive this message more than once)


To whom it may concern:


The following book, which describes the application of neural
networks to real-word data, is now available:

Ralf Kompe
Prosody in Speech Understanding Systems
Lecture Notes in Artificial Intelligence, Vol. 1307
Subseries of Lecture Notes in Computer Science
Springer
Berlin, New York
1997
(370 pages)
ISBN 3-540-63580-7



------------------

                           ABSTRACT

Prosody  covers  acoustic phenomena of speech which are not  spe-
cific to phonemes.  These are  mainly intonation, indicators  for
phrase  boundaries,  and accentuation.  This information can sup-
port the intelligibility of speech or even sometimes disambiguate
the meaning.

The  aim  of this book is to describe algorithms developed by the
author for the use of prosodic information   on  many  levels  of
speech  understanding  such  as  syntax,  semantics,  dialog, and
translation.  An  implementation of these   algorithms  has  suc-
cessfully  been  integrated into the speech-to-speech translation
system Verbmobil  and in the dialog system Evar. This is  for the
first  time that prosody is  used in  a fully  operational speech
understanding  and translation  system.  The  Verbmobil prototype
system has been  publicly demonstrated at several conferences and
industrial fairs.

The  emphasis  of the book  lies on the improvement of parsing of
spontaneous speech with the help of prosodic clause boundary  in-
formation.   Prosody  reduces  the  parse-time of word hypotheses
graphs by 92% and the number of  parse  trees  by  96%.  This  is
achieved  by integrating several knowledge sources such as proba-
bilities for prosodic events  computed by neural networks and  n-
gramms in an A*-search for the optimal parse. Without prosody the
automatic   interpretation   of   spontaneous   speech  would  be
infeasible.

The  book  gives  a  comprehensive review of the mathematical and
computational background of the algorithms and statistical models
useful   for the integration of prosody  in speech understanding.
It also shows unconventional applications of hidden  Markov  mod-
els, stochastic language models, and neural networks. The latter,
for example,  are apart from several  classification  tasks  used
for  the inverse filtering of speech signals.   The book also ex-
plains in detail the acoustic-prosodic phenomena  of  speech  and
their functional role in communication. In contrast to many other
reports, it gives a lot of examples  taken from real  human-human
dialogs; many examples are supported by speech signals accessible
over the WWW. The use of prosodic information relies on  the  ro-
bust  extraction  of relevant features from digitized speech sig-
nals, on adequate labeling of large  speech databases for  train-
ing  classifiers,   and on  the detection of prosodic events; the
methods used in Verbmobil  and Evar  are summarized  as  well  in
this  book.   Furthermore,  an overview of these state-of-the-art
speech understanding systems is given.



------------------



The book has been awarded with the "Dissertation Price" of the
German Institutes for Artificial Intelligence.




Sincerely yours,

Ralf Kompe








__________________________________________________________________________

Ralf Kompe  

Sony International (Europe) GmbH                           . . o o O O
European Research and Development Stuttgart (ERDS)         . . o o O
Advanced Developments                                      . . o o O
Stuttgarter Str. 106                                       . . o o O O
D-70736 Fellbach                                           . . o o O O O
Germany                                                    . . o o O O O
                                                           . . o o O O
Phone:  +49-711-5858-366
Fax:    +49-711-58-31-85
E-mail: kompe at fb.sony.de
__________________________________________________________________________