Content uploaded by Silvia Pfeiffer
Author content
All content in this area was uploaded by Silvia Pfeiffer
Content may be subject to copyright.
Audio Meta Data Transcription from Meeting
Transcripts for the Continuous Media Web
Claudia Schremmer1, Steve Cassidy2, and Silvia Pfeiffer1
1CSIRO ICT–Centre, Locked Bag 17, North Ryde, NSW 1670, Australia
2Macquarie University, Department of Computing, North Ryde, NSW 2109, Australia
Correspondence should be addressed to Claudia Schremmer (claudia.schremmer@csiro.au)
ABSTRACT
The Continuous Media Web (CMWeb) integrates time–continuous media into the searching, linking, and
browsing functionality of the World Wide Web. The file format underlying the CMWeb technology, Annodex,
streams the media content multiplexed with XML–markup in the Continuous Media Markup Language
(CMML). CMML contains information relevant to the whole media file (e.g., title, author, language) as well
as time–sensitive information (e.g., topics, speakers, time–sensitive hyperlinks). This paper discusses the
challenges of automatically generating Annodex streams from complex annotated recordings collected for use
in linguistic research. We are particularly interested in annotated recordings of meetings and teleconferences
and regard Annodex and its media browsing paradigm as a novel and rich way of interacting with such
recordings. The paper presents our experiments with generating CMML and their corresponding Annodex
files from hand annotated meeting recordings.
1.INTRODUCTION
In mid–2003, CSIRO launched the Continuous Me-
dia Web (CMWeb) [1, 13, 12], designed to solve the
problem of “dark matter” of time–continuous media
such as audio and video on the Web. The motiva-
tion for the proposed streamable annotated and in-
dexed Annodex file format is the integration of time–
continuous media into the URL–based hyperlink-
ing paradigm, resulting in “surfable” and “search-
able” media. In this new proposed standard which
has been submitted to the IETF for standardization
[10], users can not only hyperlink to, e.g., an au-
dio file, but into a specific time interval containing
the information sought–after [9]. The audio itself
is annotated with HTML–like markup information
and might link to yet other resources on the Web
(e.g., text, audio, images, video), enabling the user
to switch from one digital resource to another, just
like “browsing” a Web site. The technology is based
on the observation that the data stream of existing
time–continuous data can be subdivided into seg-
ments based on a semantic concept such as topic
boundaries, and that this structure enables access
to interesting subparts or clips of the stream.
One significant source of annotated time–continuous
media is that of linguistic corpora. These are col-
lections of language data, often speech, audio and
video, which have been annotated for their linguistic
content. The annotations range from simple time–
aligned transcriptions to complex structures of en-
coded dialogue including forward and backward re-
lations amongst turns, annotations of overlapping
speech etc. The speech recognition community is
now targeting meeting recordings. The term “meet-
ing” is broad, embracing face–to–face meetings as
well as teleconferences. They range from formal pre-
sentations over organized group meetings to informal
“brain storming” sessions. In all cases, the nature
of the recording is that the number of participants
is not restricted (and often not known) and the lan-
guage used is informal and spontaneous.
AES 25TH INTERNATIONAL CONFERENCE, LONDON, UNITED KINGDOM, 2004 JUNE 17–19
1
SCHREMMER ET AL. AUDIO META DATA FOR THE CMWEB
This paper describes some experiments in transform-
ing meeting transcriptions into the streaming me-
dia format Annodex of the Continuous Media Web.
This is motivated in part by the need to automate
the creation of Annodex content: Speech is a rich
source of content for media data. The experiments
also form an additional part of the validation of the
CMWeb standards for streaming and linking time–
continuous media.
Section 2 introduces the Continuous Media Markup
Language (CMML) that allows authoring of tex-
tual markup for time–continuous data in Annodex
format. Section 3 outlines the concept of linguis-
tic annotations for meeting recordings. Section 4
presents the core of this paper and describes the
means of transcribing complex linguistic transcripts
into CMML. In Section 5, we summarize the bene-
fits from our approach. Some future challenges are
discussed in Section 6.
2.CONTINUOUS MEDIA MARKUP LAN-
GUAGE
The Annodex file format [1, 10] for the Continu-
ous Media Web contains a media data stream inter-
spersed with islands of XML markup that contains
annotations and supports hyperlinking to other clips
and Web resources. Web server extensions allow
search engines to retrieve the XML markup repre-
senting the Annodex file. For producing an Annodex
file, a markup file written in the Continuous Media
Markup Language CMML [11] is interleaved with
the media stream. Figure 1 illustrates the merging
process. Here, anxenc is short for Annodex encode.
2.1.CMML Syntax
A sample CMML file is given below. The XML
markup of a head tag embraces information for the
complete media document. It contains structured
textual annotations in meta tags as well as unstruc-
tured textual annotations in the title tag. Struc-
tured annotations are name-value pairs which may
follow a new or an existing meta data annotation
scheme (such as the Dublin Core [6] used in the
example below). Following the head element are
one or more clips, i.e., semantic units of the me-
dia file, which might contain meta data, hyperlinks,
active
CMML file
time−continuous data
merger
(anxenc)
head clip clip clip clip
active active Annodex
time
time
Fig. 1: Merging XML markup in CMML format with
a media file.
keyframe images, and textual descriptions for a tem-
poral segment of the media document. The clips
are to be interleaved time–synchronously into the
Annodex bitstream.
<?xml version="1.0" encoding="UTF-8" ?>
<!DOCTYPE cmml SYSTEM "cmml.dtd">
<cmml>
<stream timebase="npt:0">
<import mimetype="audio/wav"
src="projectmeeting3.wav"
start="npt:0"/>
</stream>
<head>
<title>
Project Meeting Nov 2003
</title>
<meta name="DC.author" content="CSIRO"/>
<meta name="DC.date.created"
content="2003/11/10"/>
<meta name="DC.description"
content="Meeting Recording"/>
<meta name="DC.format" content="audio"/>
<meta name="DC.language" content="en"/>
[...]
</head>
<clip id="agenda" start="npt:0">
<a href="http://foo.bar/agenda.htm">
Agenda of the meeting on 2003/11/10
</a>
<img src="http://foo.bar/chair.jpg"/>
<desc>
AES 25TH INTERNATIONAL CONFERENCE, LONDON, UNITED KINGDOM, 2004 JUNE 17–19
2
SCHREMMER ET AL. AUDIO META DATA FOR THE CMWEB
Peter outlines the agenda for the meeting.
It includes...
</desc>
</clip>
<clip id="topic1" start="npt:230" end="npt:423">
<a href="http://foo.bar/topic1.htm">
Hyperlink to the HTML pages of the
meeting slides, topic 1.
</a>
<desc>
Topic 1: How can we make a meeting more
efficient? There are several options...
</desc>
</clip>
<clip id="topic2" start="npt:430">
[...]
</clip>
</cmml>
As the sample file shows, the structure of the CMML
annotations is fairly simple. A time–continuous
media file is subdivided into clips defined by a
start time and an end time. The CMML DTD
(http://www.annodex.net/DTD/cmml.dtd) allows
most of the tags and parameters (such as the end
parameter in the clip tag) to be optional. Each
clip might contain amongst others a hyperlink <a>
to a related Web resource, an image <img> repre-
senting the content of the clip, a free-text description
<desc> of its content, and a set of meta data <meta>
providing structured annotations for the clip.
The user experience of consuming an Annodex me-
dia file is such that while, e.g., listening to an au-
dio file, the annotations and links change over time:
During the playback of a recorded speech file, the
islands of CMML markup data that are defined in
the clip tags and that are interspersed with the
media file itself, enter and quit their active periods
through their start and optional end tags. If at
any arbitrary offset into the speech file, there is a
clip active for that period, its markup is extracted.
If the clip under consideration happens to have a
hyperlink defined through its <a> tag, the user may
decide to activate it during the duration of this clip.
2.2.Multitrack Paradigm
So far, we introduced the basic structure of CMML.
For the break–down of linguistic transcripts (see Sec-
tion 3) into the rather linear consumption structure
of CMML for Annodex, we make use of CMML’s
multitrack paradigm, which enables a set of tem-
porally overlapping annotation tracks per Annodex
file. This is implemented through the definition of
atrack attribute within a clip tag. If no track is
set, it defaults to track="default". See Figure 2
for how a time–continuous media file is interspersed
with islands of CMML annotations that belong to
different annotation tracks. Whereas clips within
the same annotation track cannot overlap, the active
period for the markup of clips belonging to different
tracks might overlap in time.
merger
(anxenc)
CMML markup
track 1 track 2 track 3
time−continuous data
track 1
track 2
track 3
media data
Fig. 2: The multitrack paradigm of CMML: A media
file can be annotated with an arbitrary number of
annotation tracks. The resulting Annodex bitstream
may contain more than one active clip for each given
point in time.
The multitrack paradigm allows the consumer of a
meeting recording to change his viewing angle. If,
for example, there are three different topic tracks to
a recording, the user can choose which topic he is
interested in and only use the annotations related
to his choice. For our speech annotations, we have
experimented with different tracks according to the
speaker. If, for example, we have a speaker peter,
the clip syntax as described in Section 2.1 is en-
hanced by the track="peter" attribute:
<clip id="id25" track="peter"
start="102.300000" end="104.310000">
AES 25TH INTERNATIONAL CONFERENCE, LONDON, UNITED KINGDOM, 2004 JUNE 17–19
3
SCHREMMER ET AL. AUDIO META DATA FOR THE CMWEB
The realization of how this is implemented in the
user interface is left to the application, i.e., to the
respective Annodex browser. The generic implemen-
tation allows the user to select his annotation track
of interest from a list. A specific implementation for,
e.g., multispeaker annotation tracks could also rep-
resent all the speakers and their speech in different
sections of the screen in parallel.
3.LINGUISTIC ANNOTATION
Linguistic data is annotated in many ways and in-
cludes complex interlinking between elements. In
particular, annotations can often involve multiple
overlapping hierarchies: For example describing two
simultaneous speakers, two concurrent topics for one
piece of spoken language, or alternate structural
views of a piece of language data. Hence, the un-
derlying data structure of linguistic annotations is
highly complex. One of the goals of recent work in
the area of linguistic annotations has been to define
a standard data model able to represent all the var-
ious kinds of annotation: From simple time–aligned
transcriptions to complex structures including for-
ward and backward relations amongst turns, anno-
tations of overlapping speech etc. One such proposal
is the Annotation Graph [2] which represents anno-
tations as a directed acyclic graph. Some of our
recent work in this area has been trying to identify
a sufficiently powerful query language for use on an-
notations [3, 4].
For this work, we have used linguistic transcriptions
in the Universal Transcription Format UTF [15].
The goal of this linguistic annotation format is to
capture the orthography of spoken words in record-
ings of speech and can also include annotations
which associate certain signal, speaker, and con-
tent conditions with the speech and its transcription.
The convention is formatted using SGML markup.
SGML tags identify the location of the speech within
the recorded waveform using start/end time at-
tributes, and they identify the corresponding tran-
scription by enclosing it within the span of these
tags. UTF has at its core a notion of speaker turns.
A speaker turn marks the extent and content of the
words spoken by a single speaker, spoken in a par-
ticular speaking style [15].
There are four categories of tags in a UTF doc-
ument: structural, state change, pseudo bracket-
ing, and lexical. The pseudo bracketing tags do
not use the SGML hierarchy of begin and end tags,
but instead explicit begin tags and end tags (e.g.
<b_unclear>unclear speech<e_unclear>).
In the following, we use hand annotated meeting
transcripts in UTF format developed in the ICSI
Meeting Room Project [7] that were used in the
NIST Rich Transcription Evaluation Spring 2004 [8].
We will use the following excerpt of the ICSI meeting
transcription ICSI_20010208-1430.utf to monitor
its changes throughout our transcoding process dur-
ing the remain of this paper.
<utf dtd_version="utf-1.0" audio_filename="ICSI_
20010208-1430" scribe="dummy" language="English"
version="20" version_date="08-Aug-2001">
<conversation_trans recording_date="unknown">
[...]
<turn startTime="101.008000" endTime="102.383000"
speaker="me011" spkrType="male" channel="5">
but that <contraction e_form="[wo=>will][n’t=>
not]">won’t take too long .
</turn>
<turn startTime="101.040000" endTime="101.510000"
speaker="me013" spkrType="male" channel="2">
Right .
</turn>
<turn startTime="102.300000" endTime="104.310000"
speaker="me013" spkrType="male" channel="2">
<acronym>O<acronym>K , agenda items ,
</turn>
<turn startTime="104.750000" endTime="106.241000"
speaker="me013" spkrType="male" channel="2">
<nonlexeme>Uh , we have digits ,
</turn>
[...]
</conversation_trans>
</utf>
4.TRANSCRIPTION INTO CMML
The scope envisaged in this work is to define struc-
tures within linguistic annotations that are hierar-
chical, thus allowing XML concepts and tools to ma-
nipulate linguistic data. This work has now lead
to an investigation of the use of XSLT stylesheets
AES 25TH INTERNATIONAL CONFERENCE, LONDON, UNITED KINGDOM, 2004 JUNE 17–19
4
SCHREMMER ET AL. AUDIO META DATA FOR THE CMWEB
[5] for transforming (hand) annotated meeting tran-
scriptions into display formats like CMML.
This section describes our work to automatically
generate a CMML file for the Annodex file format
from annotated meeting transcripts in the UTF for-
mat. The advantage of using these speech tran-
scripts is that there is a large number available.
Hence, it allows us to concentrate on the high–level
problems of how to re–purpose existing material for
our specific needs.
4.1.XML–ification of UTF
As we have seen in Section 3, linguistic annotations
in the UTF format are not strictly XML. In order to
be able to engage XML tools to retrieve the interest-
ing information, we modified the original transcripts
ICSI_date-time.utf as follows:
•<b_noscore ATT> has the corresponding clos-
ing tag <e_noscore>. This was modified to
<b_noscore ATT />.
•<b_unclear>unclear speech<e_unclear>
was modified to <b_unclear>unclear
speech</b_unclear>.
•<contraction ATT> was modified to
<contraction ATT />.
•<acronym>O<acronym>K was modified to
<acronym>O</acronym><acronym>K</acronym>.
•<nonspeech>swallow was modified to
<nonspeech>swallow</nonspeech>.
•...and similar modifications...
On our sample file ICSI_20010208-1430.utf, this
has the following effect:
<utf dtd_version="utf-1.0" audio_filename="ICSI_
20010208-1430" scribe="dummy" language="English"
version="20" version_date="08-Aug-2001">
<conversation_trans recording_date="unknown">
[...]
<turn startTime="101.008000" endTime="102.383000"
speaker="me011" spkrType="male" channel="5">
but that <contraction e_form="[wo=>will]
[n’t=>not]" />won’t take too long .
</turn>
<turn startTime="101.040000" endTime="101.510000"
speaker="me013" spkrType="male" channel="2">
Right .
</turn>
<turn startTime="102.300000" endTime="104.310000"
speaker="me013" spkrType="male" channel="2">
<acronym>O</acronym><acronym>K</acronym> ,
agenda items ,
</turn>
<turn startTime="104.750000" endTime="106.241000"
speaker="me013" spkrType="male" channel="2">
<nonlexeme>Uh</nonlexeme> , we have digits ,
</turn>
[...]
</conversation_trans>
</utf>
4.2.XSLT
We used an XSLT stylesheet [5] to transform the
(modified) UTF files of hand annotated linguistic
transcripts into CMML markup.
4.2.1.CMML’s head Tag
Since both UTF and CMML markup contain much
meta information, we were able to retrieve most of
the meta information necessary in the CMML head
tag directly from the UTF file. The XSLT code ex-
cerpt below produces the CMML head tag with its
corresponding meta data that are valid for the com-
plete media file.
<xsl:template match="utf">
<head>
<title>
<xsl:choose>
<xsl:when test="conversation_trans">
<xsl:text>Transcription of Meeting Recording:
</xsl:text>
<xsl:value-of select="@audio_filename" />
</xsl:when>
<xsl:otherwise>
<xsl:text>The source file is not a Meeting
Recording in utf format</xsl:text>
</xsl:otherwise>
</xsl:choose>
</title>
<meta>
<xsl:attribute name="name">
<xsl:text>DC.DESCRIPTION</xsl:text>
</xsl:attribute>
<xsl:attribute name="content">
AES 25TH INTERNATIONAL CONFERENCE, LONDON, UNITED KINGDOM, 2004 JUNE 17–19
5
SCHREMMER ET AL. AUDIO META DATA FOR THE CMWEB
<xsl:text>Automatically transformed linguistic
transcription of Meeting recording </xsl:text>
<xsl:value-of select="@audio_filename" />
<xsl:text> into CMML by an xslt script.
</xsl:text>
</xsl:attribute>
</meta>
<meta>
<xsl:attribute name="name">
<xsl:text>DATE.TRANSCRIBED</xsl:text>
</xsl:attribute>
<xsl:attribute name="content">
<xsl:value-of select="@version_date" />
</xsl:attribute>
</meta>
[...]
</head>
</xsl:template>
The result of this script run on our sample file
ICSI_20010208-1430.utf is the following CMML
code
<head>
<title>Transcription of Meeting Recording:
ICSI_20010208-1430
</title>
<meta name="DC.DESCRIPTION" content=
"Automatically [...] ICSI_20010208-1430
into CMML by an xslt script."/>
<meta name="DATE.TRANSCRIBED" content=
"08-Aug-2001"/>
[...]
</head>
4.2.2.CMML’s clip Tags
The following code excerpt transforms turn tags into
clip tags with their respective attributes.
<xsl:template match="turn">
<clip>
<xsl:attribute name="id">
<xsl:value-of select="generate-id(.)" />
</xsl:attribute>
<xsl:attribute name="track">
<xsl:value-of select="@speaker" />
</xsl:attribute>
<xsl:attribute name="start">
<xsl:value-of select="@startTime" />
</xsl:attribute>
<xsl:attribute name="end">
<xsl:value-of select="@endTime" />
</xsl:attribute>
[...]
</clip>
</xsl:template>
In each clip, we are most interested in the content of
the free–text description tags <desc>[...]</desc>
since their content is searched and indexed by Web
search engines. Due to the nature of the UTF source
files, we did not look for key frame images for the
<img> tags or for time–sensitive hyperlinks in the
<a> tags although this is an interesting future re-
search area.
The result of the above transformation is a subdivi-
sion into clips that span only a very short timeline
— usually about 0.5 to 3 seconds duration, depend-
ing on the UTF source file. With the concept in
CMML being that a clip is a semantic subdivision,
we aimed at merging consecutive clips if they belong
to the same track, i.e., speaker. When the following
test on the speaker relative to the following speaker
is true,
<xsl:when test="@speaker=substring(following-
sibling::turn/@speaker,1)">
[...]
<\xsl:when>
the clips are merged. Hence, a series of consecutive
clips with the same track attribute, like the last
three clips in our standard excerpt:
<clip id="id26" track="me011"
start="101.008000" end="102.383000">
<desc>
but that won’t take too long .
</desc>
</clip>
<clip id="id27" track="me013"
start="101.040000" end="101.510000">
<desc>
Right .
</desc>
</clip>
<clip id="id28" track="me013"
start="102.300000" end="104.310000">
<desc>
AES 25TH INTERNATIONAL CONFERENCE, LONDON, UNITED KINGDOM, 2004 JUNE 17–19
6
SCHREMMER ET AL. AUDIO META DATA FOR THE CMWEB
OK , agenda items .
</desc>
</clip>
<clip id="id29" track="me013"
start="104.750000" end="106.241000">
<desc>
Uh , we have digits .
</desc>
</clip>
result in the more compact clip presentation:
<clip id="id26" track="me011"
start="101.008000" end="102.383000">
<desc>
but that won’t take too long .
</desc>
</clip>
<clip id="id27" track="me013"
start="101.040000" end="106.241000">
<desc>
Right . OK , agenda items . Uh , we
have digits .
</desc>
</clip>
For this merging, not only two consecutive turns
have to have the same speaker attribute, but a time
threshold between the endTime tag of the previous
and the startTime tag of the consecutive turn ap-
plies. This threshold was set to 2.5 seconds based
on an indication that a break longer than this time
period often creates a semantic break in the flow of
the speech.
5.THE USER EXPERIENCE
What have we achieved so far? We started out with
a time–continuous audio file of a meeting recording
and its corresponding hand annotated file of linguis-
tic transcriptions. Meetings are informal and spon-
taneous, and the meeting transcriptions reflect this.
They contain many details of linguistic interest like:
•Ther<fragment /> there
•Why don’t you summarize the
<!--PhrasalFragment-->
•We <mispronounced /><!--prolly-->probably
[...]
•<nonspeech>swallow</nonspeech>
Our goal within the Continuous Media Web is to
establish a file format that is based on the time–
continuous media plus meta information to cater for
the needs of a Web with its searching and surfing
paradigm. The focus is less on linguistic analysis of
the recorded speech but rather to provide the user
with a pleasant consuming experience of the under-
lying media file.
The stress in the results of the previous section shall
be posed on the fact that a clip annotation as above
contains a lot of information that we are interested
in: track, start, end attributes. But most im-
portantly, it also contains free-text annotation in the
<desc> tags. Recalling that the CMWeb technology
has been built to allow Web search engines to crawl
and index this meta information, the Annodex file
resulting from the merger of CMML and the speech
file is fully searchable. In our example above, a (tex-
tual) search for “agenda items” would link exactly
into our speech segment with start="101.040000"
and end="106.241000".
Hence, we have targeted the following challenges:
•Strip the linguistic annotation from those only
of linguistic interest (by not processing them
with the XSLT stylesheet).
•Make use of the track paradigm of CMML to
break down the linear recording with its non–
linear linguistic annotation into a meaningful
semantic partition.
•Subdivide the speech recording into semanti-
cally meaningful clips.
6.FUTURE WORK
For this work, we have focused on subdividing a
CMML file into multiple tracks according to the
speaker. An interesting issue is to explore the
CMML’s multitrack paradigm for various topics.
This leads into research of how to define and au-
tomatically extract topics from a meeting recording.
Furthermore, we will investigate into transcriptions
from Annotation Graphs [2] into CMML.
AES 25TH INTERNATIONAL CONFERENCE, LONDON, UNITED KINGDOM, 2004 JUNE 17–19
7
SCHREMMER ET AL. AUDIO META DATA FOR THE CMWEB
Thirdly, it is not uncommon for a speaker in a meet-
ing to refer to a previous agreement (“As we have
agreed last week in topic 4,...”) or a document
(“Have a look at page 3 of this presentation...”).
Our vision is to use speech processing and language
technology to identify such references. This could be
used to automatically create hyperlinks from meet-
ing recordings to other digital media on the Web to
truly create a Continuous Media Web.
7.REFERENCES
[1] annodex.net. Annodex: Open Standards for
Annotating and Indexing Networked Media.
http://www.annodex.net, 2003.
[2] Steven Bird and Mark Liberman. A Formal
Framework for Linguistics Annotation. speech-
com, 2000.
[3] Steve Cassidy. Generalising XPath for Directed
Graphs. In Proceedings of Extreme Markup
Languages, Montreal, Canada, August 2003.
[4] Steve Cassidy and Steven Bird. Querying
Databases of Annotated Speech. In Maria E.
Orlowska, editor, Proceedings of the 11th Aus-
tralasian Database Conference, volume 22,
pages 12–20, 2000.
[5] World Wide Web Consortium. XSL Trans-
formations (XSLT) Version 1.0. http://www.
w3.org/TR/xslt/.
[6] Dublin Core Metadata Initiative. Dublin
Core Metadata Element Set, Version 1.1.
http://dublincore.org/documents/2003/02/04/
dces, February 2003.
[7] Nelson Morgan, Don Baron, Jane Edwards,
Dan Ellis, David Gelbart, Adam Janin, Thilo
Pfau, Elizabeth Shriberg, and Andreas Stolcke.
The Meeting Project at ICSI. In Proceedings of
Human Language Technologies Conference, San
Diego, CAL, USA, March 2001.
[8] NIST. Rich Transcription 2004 Spring Meet-
ing Recognition Evaluation. http://www.nist.
gov/speech/tests/rt/rt2004/ spring/.
[9] Silvia Pfeiffer, Conrad Parker, and Andr´e
Pang. Specifying Time Intervals in URI
Queries and Fragments of Time–Based
Web Resources (BCP) (work in progress).
http://www.ietf.org/internet-drafts/draft-
pfeiffer-temporal-fragments-02.txt, December
2003.
[10] Silvia Pfeiffer, Conrad Parker, and Andr´e Pang.
The Annodex Annotation Format for Time–
Continuous Bitstreams, Version 2.0 (work
in progress). http://www.ietf.org/internet-
drafts/draft-pfeiffer-annodex-01.txt, December
2003.
[11] Silvia Pfeiffer, Conrad Parker, and Andr´e
Pang. The Continuous Media Markup Lan-
guage (CMML), Version 2.0 (work in progress).
http://www.ietf.org/internet-drafts/draft-
pfeiffer-cmml-01.txt, December 2003.
[12] Silvia Pfeiffer, Conrad Parker, and Andr´e
Pang. Managing Multimedia Semantics, Chap-
ter: Continuous Media Web: Hyperlinking,
Search & Retrieval of Time–Continuous Data
on the Web. Idea Group, Inc., 2004. [accepted
for publication].
[13] Silvia Pfeiffer, Conrad Parker, and Claudia
Schremmer. Annodex: A Simple Architecture
to Enable Hyperlinking, Search & Retrieval of
Time–Continuous Data on the Web. In Proc.
5th ACM SIGMM International Workshop on
Multimedia Information Retrieval (MIR), pages
87–93, Berkeley, CA, USA, November 2003.
ACM.
[14] Silvia Pfeiffer and Claudia Schremmer. Multi-
Media Information Retrieval: metodologie ed
esperienze internazionali di content–based re-
trieval per l’informazione e la documentazione,
Chapter: Automated Annodexing of Meeting
Recordings. Associazione Italiana per la Doc-
umentazione Avanzata (AIDA), Roma 2004,
Italy, March 2004.
[15] A Universal Transcription Format (UTF)
Annotation Specification for Evaluation
of Spoken Language Technology Corpora.
http://www.nist.gov/speech/tests/bnr/bnews
_99/utf-1.0-v2.ps, 1998. version 1.0.
AES 25TH INTERNATIONAL CONFERENCE, LONDON, UNITED KINGDOM, 2004 JUNE 17–19
8
