Frank Enos
frank@cs.columbia.edu
20120311

This readme describes the audio files and associated transciptions, segmentations, and labelings of the Columbia Deception Corpus.

For experiment-specific terminology and definitions, please refer to my dissertation.

See /proj/speech/corpora/CDC/sub_stats_unix.txt for subect gender and interview length.

See /proj/speech/projects/deception/columbia/README.features.txt for notes on existing feature sets and their manipulation.

==================
File Convetions
==================

Each subject directory contains the same set of files; in what appears below, I have taken Subject 10B as an example.

For all subjects, left channel, denoted by _L_ in the filename, is the interviewer.  The right channel, _R_, is the subject.

Note the following conventions for Global Lie labeling:

Global Lie valence and version of the pre-interview task for the given section appears before the colon (E.g. "T/H"; the section name appears after the colon (e.g. "INTERACTIVE").

Global Lie valence is indicated as:
+ T == Truth
+ LU == Lie Up (Subject claims better task performance than was actually acheived.)
+ LD == Lie Down (Subject claims worse task performance than was actually acheived.)

Task version is indicated as:
H == Hard
E == Easy

Sections are labeled by section name.

For example, T/H:INTERACTIVE indicates that the subject is telling the truth based on having performed the "hard" version of the Interactive pre-task.


================
.wav
================

S-10B_L_16k.wav
S-10B_R_16k.wav
S-10B_LR_16k.wav

Interviewer, subject, and combined stereo audio files, respectively.


================
.TextGrid
================
S-10B_L_16k.punc.TextGrid : Interviewer transcription with punctuation.
S-10B_R_16k.punc.TextGrid : Subject transcription with punctuation.

S-10B_BigLies.TextGrid : Indicates the topic and Global Lie valence of (typically long) intervals of the interview.
S-10B_BigLiesSU.TextGrid : Indicates the topic and Global Lie valence of interview Slash Units.
S-10B_BigLiePhrases.TextGrid : Indicates the topic and Global Lie valence of interview phrases (breath delimited).


S-10B_pedal_hand_corr.TextGrid : Indicates Local Lie valence of segments of the interview; based on hand-corrected alignments of subject pedal presses.




==================
.trs
==================
S-10B_L_16k.punc.trs
S-10B_R_16k.punc.trs

Interviewer and subject versions of LDC Transcriber files used in producing hand transcriptions (correspond to the .punc.TextGrid files).

==================
.ltf
==================

S-10B.ltf

"Lie Tracker File"
File produced by the pedal-press reading Java program I wrote to record subject pedal presses; used to produce the Local Lie .TextGrid labels.


==================
.ctm / .stm
==================
(Note that these were produced by the SRI team; I never made direct use of these files.)

S-10B_L_16k-wrd.ctm
S-10B_R_16k-wrd.ctm

Word based segmentations.


S-10B_L_16k-wrd.stm
S-10B_R_16k-wrd.stm

Word based segmentations explicitly indicating inter-segment gaps.

S-10B_L_16k-phn.stm
S-10B_R_16k-phn.stm
Phone-based segmentations explicitly indicating inter-segment gaps.