BAS - General Information
The Bavarian Archive for Speech Signals (BAS)
is a public institution hosted by the University of Munich. This
institution was founded with the aim of making corpora of
current spoken German available to both the basic research and
the speech technology communities via a maximally comprehensive
digital speech-signal database. The speech material will be
structured in a manner allowing flexible and precise access,
with acoustic-phonetic and linguistic-phonetic evaluation
forming an integral part of it.
Tasks
The last few years have seen an abrupt increase in the demand
for large speech-signal data collections, both on the part of
academic investigators carrying out basic research as well as on
the part of engineers from industry working in the new
integrated field of speech and information technology. There are
many reasons for this. Primarily, however, the sudden increase
in demand must be attributed to the breakneck pace of hardware
and software development in speech signal processing. The
increasing number of techniques for acoustic-phonetic signal
processing, and the increasing amount of speech data that can be
efficiently handled and processed together generate an
accompanying demand not only for linguistically interesting text
material (which of course emerges automatically from the modern
printing industry) but also for reliably acquired and
phonetically evaluated spoken language material. A number of
national and international initiatives (such as BDSON, PHONDAT,
LDC, SPEX or COCOSDA) have, it
is true, already resulted in the collection and distribution of
large speech corpora. However,
they exhibit a variety of
formats, corresponding to the variety in the aims pursued. For
German, a central institution was clearly lacking that could
carry out such tasks within a long-term perspective. BAS will
be
responsible in Germany for these tasks for distributable corpora
of spoken German, collecting, maintaining and making them
available in standardized form.
In addition, BAS will develop its own procedures for
automatic
labelling and segmentation, making the results available with
the distributed speech corpora.
BAS was entrusted by the Bundesmministerium
für Bildung, Wissenschaft,
Forschung
und Technologie (BMBF) with the task of maintaining
both existing and future databases set up within funded
projects by the BMBF, and of exporting them (after any restrictions on
availability have expired) within the EU as well as to the
Linguistic Data Consortium (LDC). Imported databases are to be
converted by BAS to a
standardized form, enabling them to be expoited in all BMBF-
funded speech projects for a fraction of the cost and effort
usually incurred.
Aims
The first aim of BAS will be to satisfy the immediate demand
for
spoken language data recorded under controlled conditions of the
kind required for speech technology development in German. This
will include development of new techniques for efficient
handling of and access to very large quantities of phonetic
data, independent of the location and the nature of the storage.
In addition to typical application-oriented corpora such as
Polyphone this first aim will concentrate on establishing
a representative database of publically spoken German.
The second goal consists in the long term
development of a (more or less)
Complete Phonetic
Theory \mbox{(CPT) of spoken German. For this endeavour, the central
category will no longer be the speech sound but rather
the word as the lexically given unit. The great variability
characterizing the pronunciation of words in running speech as
opposed to citation form will be systematically documented and
related to the communicative information content.
External cooperations
The Leibniz Rechenzentrum
München (LRZ) -- which is connected to the site via fiber
optic
data link -- provides the Archive with mass storage and network
support within the framework of the TERABACK project.
The BAS is
keen to cooperate with all institutions in the German speaking
area interested in contributing to the common goal.
Most of the projects will be financed by interested partners in
industry, by public grants or by European projects.
The BAS produces speech ressources either by public
funding or industrial cooperations. Speech ressources funded
exclusively by
public money are available without restrictions immediately after the
release for everybody. Industrial partners that have significantly
contributed to the
production of the ressource are granted a period of one year after the
release to exploit the data exclusively.
After that period the ressource is distributed via the BAS
either unrestricted or under license.
Staff
Christoph Draxler studied Computer Science at the Technical
University of Munich, Germany. He earned his PhD in 1991 from the
University of Zurich, Switzerland in the field of databases.
Since 1991 he has been working at the
Institut für Phonetik und Sprachliche Kommunikation mainly within
the PhonDat and VERBMOBIL projects.
His main interests include logical programming, databases and
multi-media applications.
Phil Hoole
Florian Schiel received his
Dipl.-Ing. and Dr.-Ing. degrees from the Technical University in Munich
in 1990 and 1993 respectively, both in electrical engineering.
Since 1993 he has been with the Institute of Phonetics, University of
Munich, participating in the VERBMOBIL project. His main interests are:
speaker adaptation, German phonetics, computational phonology,
automatic
analysis of very large speech corpora.
|