Error Region Analysis (ERA) program input file format.
Lin Chase
Carnegie Mellon University
20 July 1995

(Modified December 21 1995 to cover input phone segmentations for both
  REFerence and HYPothesis information.)

-----
<int num_utterances> UTTERANCES
UTT <utterance_id_tag1>
REF
<int num_segmentsR> SEGMENTS
<string word1R>
<int start_frame1R>
<int end_frame1R>
<int32 acoustic_score1R>
<int32 language_score1R>
<string language_score_source1R>
<int num_phone_segments1R>
<string phone1R1>
<int start_frame1R1>
<int end_frame1R1>
<int32 acoustic_score1R1>
<string phone1R2>
.
.
.
<string word2R>
.
.
.
HYP
<int num_segmentsH> SEGMENTS
<string word1H>
<int start_frame1H>
<int end_frame1H>
<int32 acoustic_score1H>
<int32 language_score1H>
<string language_score_source1H>
<int num_phone_segments1H>
<string phone1H1>
<int start_frame1H1>
<int end_frame1H1>
<int32 acoustic_score1H1>
<string phone1H2>
.
.
.
<string word2H>
.
.
.
UTT <utterance_id_tag2>
REF
.
.
.
HYP
.
.
.
EOF
-------------
Notes: 
1.  The integer in front of the token "UTTERANCES" indicates how many "UTT", "REF" and
"HYP" entries there will be in the file. 
2.  For each REF the integer in front of the token "SEGMENTS" indicates the number of 
word segmentations that should be included before the next instance of the "HYP" token
is encountered. 
3.  For each HYP the integer in front of the token "SEGMENTS" indicates the number of
word segmentations that should be included before the next instance of the "REF" token
is encountered.
4.  "language_score_source" strings can be used to indicate algorithmic origins of 
language model scores, such as the branch of the Katz backoff algorithm used.  The 
blank string "" should be used if you'd like to skip this bit.
5.  The start frames of the REF and HYP sequences must be the same.  The end frames of 
the REF and HYP sequences must be the same.  The start frame of one segment within a
REF/HYP sequence must be one integer count greater than the end frame of the previous
segment in the sequence.
6.  Phone segmentation start frames for initial phones in words must be exactly
the same as start frames of their parent word.
7.  Phone segmentation end frames for final phones in words must be exactly the
same as end frames of their parent word.
8.  Phone segmentations within words must not overlap and must completely 
partition the set of frames that make up the parent word.