Mark (.mrk) file specifications, Updated 03/16/92:

The mark files contain time-aligned word transcriptions.

1. In .mrk files, each record must have 4 fields; the first is the
talker, A or B; the second is the estimated start time of the current
word, the third is the duration of the word, and the fourth is the
word.

2. The first field may contain the dummy symbol "*" when the event in
field 4 is not attributed to either speaker.  Examples: "[beep]" at
the beginning indicating the dtmf tone, or the "..." at the end
indicating that the conversation was cut off (timed out).

  *        *       *    [Beep]
 @A       1.36    0.28  Okay,
 @A       1.64    0.08  I


3. Fields 2 and 3 may contain the dummy symbol "*" when the event in
field 4 is not a word, but a nonspeech event, a comment, or a
stand-alone string of punctuation (i.e., something which does not
receive a duration from the time alignment algorithm), as well as the
cases in (2) above.

  A          *       *  [lipsmack]
  A          *       *  {pause}
  A     311.02    0.62  economic


4. Fields 2 and 3 may contain the dummy symbol "*" when there is
simultaneous speech by A and B.  Here the time alignment algorithm is
allowed to recognize the speech of EITHER talker--usually the louder
one.  The other talker's speech is not time aligned during this
period, but the words should be attributed to A and B correctly.

  A     113.96    0.24  thing
  A     114.20    0.10  is
  A     114.30    0.44  still,
  A          *       *  {pause}
  A          *       *  #you
  A          *       *  know#
  A          *       *  --
  B     116.40    0.20  #Your
  B     116.60    0.56  education.#
  A          *       *  --
  A     117.16    0.22  getting
  A     117.38    0.10  your
  A     117.48    0.60  education.


5. In field 1, "@" or "@@" may occur before A or B to mark suggested
trim points for topicality purposes.  However, we have REMOVED ALL
DOUBLE ASTERISKS "**" from this location in both .mrk and .txt
files.  These originally signified that the two talkers' speech
overlapped during this turn, without having to indicate exactly which
words were involved.  For short and simple episodes of overlap,
transcribers were supposed to be more exact, placing pound signs "#"
around the two simultaneously spoken phrases.  See the previous
example.

This was not specific enough, either for some transcribers or for the
time alignment program, so we began enforcing the stricter standard
for all cases of simultaneous speech.  Some of the released files
still had the old asterisk notation when sent out in December.


6. In some of the more recently processed files, the start times of 
keywords found not to properly marked are preceded by "&&".  This
was done during a check of keywords after automatic marking had been
completed.  The ampersands were added if playing the keyword as
marked produced less than a full keyword or parts of other words.
The presence of silence or background noise before or after the
keyword did not cause them to be added.


7. It should be borne in mind that these markings were created by
an automatic procedure which did not model non-speech events, and
were only intended to be approximate.  Some marks, including some
of keywords, are quite incorrect and may not overlap in time the
actual word.  NIST will provide official markings of the keywords
in June 1992.