Czech Broadcast Conversation MDE Transcripts,
Linguistic Data Consortium (LDC) catalog number LDC2009T20 and ISBN 1-58563-520-0,
was prepared by researchers at the University of West Bohemia, Pilsen, Czech
Republic, and consists of approximately 33 hours of transcribed speech from
Radioforum, a talk show broadcast on Czech Radio 1. The audio files corresponding
to the transcripts in this corpus are contained in Czech
Broadcast Conversation Speech (LDC2009S02). These corpora join LDC's other
Czech broadcast data sets: Czech
Broadcast News Speech (LDC2004S01), Czech
Broadcast News Transcripts (LDC2004T01), Voice
of America (VOA) Czech Broadcast News Audio (LDC2000S89), and Voice
of America (VOA) Czech Broadcast News Transcripts (LDC2000T53).
Broadcast Conversation Speech consists of 72 single channel recordings of
Radioforum, a live talk program broadcast by Czech
Radio 1 (CRo1) every weekday evening. A total of 40 hours of recordings
were collected during the period from February 12, 2003 through June 26, 2003.
Individual recordings range from 27 minutes to 36 minutes each. Radioforum's
format consists of invited guests (most often politicians) spontaneously answering
topical questions posed by one or two interviewers. The number of interviewees
in a single program varies from one to three, but typically, one interviewer
and two interviewees appear in the program. The material includes passages of
interactive dialogue, but longer stretches of monologue-like speech comprise
the majority of the collected data. Radioforum also has an interactive segment
where listeners call the studio and ask their own questions. That telephony
speech was not transcribed in the current release.
Czech Broadcast Conversation MDE Transcripts was
created to extend Metadata Extraction (MDE) research to conversational Czech.
The goal of MDE is to take raw speech recognition output and refine it into
forms that are of more use to humans and to downstream automatic processes.
In simple terms, this means the creation of automatic transcripts that are maximally
readable. This readability might be achieved in a number of ways: removing non-content
words like filled pauses and discourse markers from the text; removing sections
of disfluent speech; and creating boundaries between natural breakpoints in
the flow of speech so that each sentence or other meaningful unit of speech
might be presented on a separate line within the resulting transcript. Natural
capitalization, punctuation and standardized spelling, plus sensible conventions
for representing speaker turns and identity are further elements in the readable
The transcripts and annotations in this corpus are stored in three different
formats: TRS (Transcriber - http://trans.sourceforge.net), QAn (Quick Annotator
- http://www.mde.zcu.cz/qan.html), and RTTM. TRS represents a standard speech
transcript. QAn and RTTM contain essentially identical information about structural metadata
(MDE); the main difference between them is formatting. Character encoding in all files is ISO-8859-2.
All filenames have the form rfYYMMDD.format where "rf" stands for
Radioforum, the following six digits indicate the date of broadcast, and the
extension ".format" corresponds to the data format of the particular
file ".trs", ".qan", or ".rttm".
More information can be found on the website Structural Metadata Annotation for Czech.
The radio programs recorded for this corpus were transcribed with two purposes.
First, in order to produce precise time-aligned verbatim transcripts of the
audio recordings, manual transcripts were created using guidelines based on
those employed in
Czech Broadcast News Transcripts (LDC2004T01). Second, the transcripts were
annotated wiith MDE markup to provide structural information about the conversations.
Manual time-aligned verbatim transcription
The original guidelines for time-aligned verbatim transcription used for the
Czech broadcast news data were adjusted to better accommodate specifics of the
recorded broadcast coversation. Those revised guidelines instructed annotators
how to deal with the following phenomena, among others:
- Speaker turns: a corresponding time stamp and speaker ID are inserted every
time there is a speaker change in the audio.
- Turn-internal breakpoints: to break up long turns, breakpoints roughly corresponding
to 'sentence' boundaries within a speaker turn are inserted.
- Overlapping speech: an overlapping speech region is recognized when more
than one speaker talks simultaneously; within this region, each speaker's
speech is transcribed separately (if intelligible).
- Background noises: [NOISE] tags are used to mark noticeable background noises.
- Speaker noises: speaker-produced noises are identified with one of the following
tags: [BREATH], [COUGH], [LAUGH], [LIP-SMACK].
- Filled pauses: filled pauses produced by a speaker to indicate hesitation
or to maintain control of a conversation are transcribed either as [EE-HESITATION]
or as [MM-HESITATION], based on their pronunciation.
- Interjections: certain interjections typically used as back channels or
to express speaker's agreement or disagreement are transcribed using
the [HM] (agreement) and [MH] (disagreement) tags.
- Unintelligible speech: regions of unintelligible speech are marked with
a special symbol.
- Numbers: all numerals are transcribed as complete words.
- Mispronounced words: mispronounced words (reading errors, slips of the tongue)
are transcribed in the spelling corresponding to their pronunciation in the
audio (i.e., the incorrect pronunciation is represented) and marked with a
- Word fragments: the pronounced part of the word is transcribed and a single
dash is used to indicate point at which word was broken on.
- Punctuation: standard punctuation (limited to commas, periods, and question
marks) is used to enhance transcript readability.
Because the verbatim transcripts were created by a large number of annotators,
they were manually revised for maximum correctness and consistency.
MDE is an annotation task which annotates Edit Disfluencies (repetitions,
revisions, restarts and complex disfluencies), Fillers (including, e.g., filled
pauses and discourse markers) and SUs, or syntactic/semantic units. Originally,
the structural MDE annotation standard was defined for English. When developing
structural metadata annotation guidelines for Czech, the
guidelines developed by LDC for English were followed to the extent possible.
Lanaguage-dependent modifications were made based on the description of the
syntax of Czech compound and complex sentences. MDE Annotation marks the following
- Edit Disfluencies: Edit disfluencies, or speech repairs,
occur when speakers correct or alter their utterances or abandon them entirely
and start over.
- Fillers: While the term filler has traditionally been synonymous
with filled pause, SimpleMDE uses the term to encompass a broad set of vocalized
space-fillers: filled pauses (FPs), discourse markers (DMs), explicit editing
terms (EETs) and asides/parentheticals (A/Ps).
- Sentence-like units: One of the goals of MDE annotation is the
identification of all units within the discourse that function to express
a complete thought or idea on the part of the speaker.Within MDE these elements
are called SUs (Syntactic, Semantic or Slash Units).
The table below contains details about the audio files and the transcripts:
|Number of shows
|Number of word tokens
|Number of unique words
|Duration of transcribed speech
|Total number of speakers
The Czech Broadcast Conversation MDE Transcripts employs three transcription formats. A sample of each is included below.
- TRS-Transcriber-Provides basic transcription of speech.
- QAN-Quick Annotator, the annotation format used to provide structural metadata.
- RTTM These annotations provide structural metadata using a format similar EARS MDE.
The completion of this corpus was facilitated by funding provided by the Ministry of Education of the Czech Republic under projects No. ME909 and 2C06020.
Portions © 2003 Cesky rozhlas 1 Radiozurnal, © 2009 Trustees of the
University of Pennsylvania