=============================================================
Levantine Arabic Conversational Telephone Speech, Transcripts
=============================================================

This database contains 982 Levantine Arabic speakers taking part in
spontaneous telephone conversations in Colloquial Levantine Arabic.  A
total of 985 conversation sides are provided (there are three
speakers who each appear in two disctinct conversations).  The average
duration per side is between 5 and 6 minutes.

This corpus was collected and transcribed in 2004 by Appen Pty Ltd (Appen), Sydney,
Australia.

Transcript Data
---------------

The "transc" directory contains three subdirectories:

 devtest --  12 two-channel audio files
 train1c --   5 single-channel audio files
 train2c -- 478 two-channel audio files

The single-channel files represent just one side of a normal
conversation.  The "devtest" set represents a relatively balanced
(representative) sample drawn from the total pool of collected calls,
based on a test-set selection process applied by the National Institute of Standards and Technology (NIST) and based on
demographic, phone and audit information as provided by Appen.

The file naming convention for the audio data is as follows:

 arb_lev-<DATE>_<TIME>-<CHANNEL>.sph   (for single-channel files)
 arb_lev-<DATE>_<TIME>.sph             (for two-channel files)

Where:  <DATE> = date of recording (YYYYMMDD)
        <TIME> = time of recording (HHMMSS, 24 hour clock)
        <CHANNEL> = "a" or "b"
        
The date and time are the time at the place of recording (i.e. Sydney,
Australia, where the recording platform was located).


Transcript Format
-----------------

Each transcript file is a flat, plain-text table, where each line
contains information for a single contiguous utterance, presented via
the following tab-delimited fields:

 1. beginning and ending time stamps, in seconds; each time stamp is
    in square brackets, and the two values are separated by a space
    (e.g. "[5.7189] [9.2135]" -- here, duration is about 3.5 sec)

 2. channel/speaker-ID ("A:" or "B:")

 3. MSA-based "consonant skeleton" orthography for the utterance,
    using Arabic script characters in UTF-8 encoding

 4. Fully "diacritized" orthography for the utterance, reflecting the
    actual pronunciation, using Arabic characters in Buckwalter
    (ASCII) transliteration

Within fields 3 and 4, word boundaries are marked by space characters
in the normal way, following common practices of Arabic orthographic
convention (e.g. all definite articles and many conjunctions and
prepositions are attached as prefixes to the following word).

Transcript tokens enclosed in single parentheses -- e.g. "(DHk)" --
represent annotation marks for non-speech events or conditions, such
as laughter, noise, etc.  Multi-token strings within single
parentheses involve words in some other language (typically English)
or some other Arabic dialect.  Other strings within single parens are
annotations about the pronunciation of a "lam" or "t-marbuta" suffix
in the preceding word; e.g. "(lam Pronounced)", or "(ap Pronounced)".

Double parentheses, either with or without tokens enclosed within them
-- e.g. "(())", "((word))" or "((word1 word2))" -- represent regions
where the transcriber was unable to tell for sure what was said.


Documentation Files
-------------------

The "doc" directory contains the following two tab-delimited table
files, which provide speaker demographics, audit results, and
phone-set information for the respective subsets of audio files/calls:

 devtest_calls.tab  --  12 rows
 training_calls.tab -- 483 rows (both single- and two-channel calls)

Both tables contain the following data fields (these column headings
are listed in the first line of each file):

 File-ID
 Date
 Time
 A-ID
 A-Region
 A-Age
 A-Gender
 A-Phone  (type)
 B-ID
 B-Region
 B-Age
 B-Gender
 B-Phone  (type)
 Topic   (assigned for conversation)
 Length  (MM:SS)

In the case of the single-channel training files, the fields
associated with the missing side are filled with "(na)".

In addition, the following two files contain word inventory and
lexicon data extracted from the transcripts:

 arb_lev_lexicon.txt
 arb_lev_reference.txt

The arb_lev_lexicon.txt file contains all of the complete Arabic
words in the transcriptions, with the MSA spelling in UTF-8, followed
by the corresponding standardised Gulf word in Buckwalter, and then
the MCA word in Buckwalter (as it appears in the pronunciation string
in the transcript files).

The arb_lev_reference.txt contains all other words in the
transcripts, such as foreign words, mispronunciations, fragments,
unintelligible words and so on; the first field is the UTF-8 form,
and the second is the Buckwalter form.