asclite - Multi-dimensional alignment program.
asclite <Input Options> [<Alignment Options>] [<Output Options>] [Scoring Report Options>]
Asclite is multi-dimensional extension of the Dynamic Programming solution to Levenshtein Edit Distance calculations capable of evaluating STT and SASTT systems during periods of overlapping, simultaneous speech. It handles different input formats: RTTM, CTM, STM, and TRN and generates several data and statistical formats: SGML, SUM, RSUM, and Alignment Information. STDOUT is used to generate reports and STDERR is used for logging purposes. The alignment information is considered as a log information.
Examples:
asclite -r example.rttm rttm -h example2.rttm rttm Example -adaptive-cost -f 6 -o sgml asclite -r example.rttm rttm -h example2.rttm rttm Example -f 0 -o sum stdout
No known bugs.
There 2 different cost models used in ASCLITE: Word-base cost and Time-base cost.
The default one is based on the words. It assigns costs regarding the action taken on the word (correct, insertion, deletion, substitution, etc.): Correct (the two word are aligned together but have not the same syntax) [cost of 0], Substitution (the two word are aligned together and have the same syntax) [cost of 400], Deletion (word that appears in the hypothesis but not in the system output) [cost of 300], Insertion (word that not appears in the hypothesis but in the system output).
If there is no distinction between hypotheses and system outputs, it uses a 'generic' cost model. The 'generic' cost model is not working on pair of words but by set of words; the set of words representing the state of the alignment matrix. It calculates the number of 'insertions' and 'deletions' and the number of words that have the same syntax. It adds the costs of all the insertions and deletions, plus the cost of correct to the maximum number of words which have the same syntax and the cost of substitution to the rest, and this to the set of words.
The second one is based on the times. It assigns costs regarding the times of the words. The cost if the sum of the differences between the begin times and end times of the words.
For example:
B E |---| is the word in the timeline. ***** is the difference. B E Word1: |-------|****** Word2: *****|--------| B E
Cost: 11 (number of *)
If there is no distinction between hypotheses and system outputs, it uses a 'generic' cost model. The 'generic' cost model, as for thw word-base model, is not working on pair of words but by set of words; the set of words representing the state of the alignment matrix. It uses the minimum and maximum times of all the words. Uses that times to calculates the costs by adding the differences between the minimum time versus the begin times, and the maximum time versus the end times. For the deletions, it just add to the cost the difference between the 2 maxima.
For example:
B E |---| is the word in the timeline. ***** is the difference. Word1: **|------------|***** Word2: *****|--------|****** Word3: **|---|************** Word4: |--------|*********** Deletion1: ********************* Word5: ************|-------| Cost: 78 (number of *)
The UEM filtering works as both a filter and a semantic checker. The times within the UEM and reference files are expected to be coherent meaning there are no UEM segment times that are within the bounds of a reference segment.
It works as a filter because it keeps and scores only the reference segments that are encapsulated by the begin/end time definitions in the UEM (Seg4 in the example). Conversely, reference segments fully outside the UEM (Seg5-6 in the example) are removed from the scoring. If there is no UEM entry defined for a particular file/channel, then the segments with the same file/channel are removed as well.
It works as a checker by detecting if some segments are not fully inside (or outside) the UEM time definitions (Seg1-3 in the example). These conditions are fatal errors.
B E |---| is the segment in the timeline.
UEM: |-------------| Seg1: |-------| [out - error] Seg2: |-------| [out - error] Seg3: |-------------------| [out - error] Seg4: |-------| [ in - ok ] Seg5: |--| [out - info ] Seg6: |---| [out - info ]
The program reports an error when a reference segment is not fully inside the UEM (Seg1-3 in the example) then exits. If a reference segment is fully outside the UEM (Seg5-6 in the example) then asclite outputs informational messages about the segment.
Where errors are found in the UEM filtering, asclite exits.
The UEM filtering works not with the begin and end time of every hypothesis segments/tokens, but rather the mid-time. If the mid-time is outside the UEM time definitions then the segment/token is removed from scoring and an information message is outputted.
[1] Multiple Dimension Levenshtein Edit Distance Calculations for Evaluating Automatic Speech Recognition Systems During Simultaneous Speech, Jonathan Fiscus, Jerome Ajot, Nicolas Radde and Christophe Laprun, Proceedings of Language Resources and Evaluation (LREC), Genoa, Italy, May 2006.
Public Domain
This software was developed at the National Institute of Standards and Technology by employees of the Federal Government in the course of their official duties. Pursuant to title 17 Section 105 of the United States Code this software is not subject to copyright protection and is in the public domain. ASCLITE is an experimental system. NIST assumes no responsibility whatsoever for its use by other parties, and makes no guarantees, expressed or implied, about its quality, reliability, or any other characteristic. We would appreciate acknowledgement if the software is used.
THIS SOFTWARE IS PROVIDED ``AS IS.'' With regard to this software, NIST MAKES NO EXPRESS OR IMPLIED WARRANTY AS TO ANY MATTER WHATSOEVER, INCLUDING MERCHANTABILITY, OR FITNESS FOR A PARTICULAR PURPOSE.