NAME

SYNOPSIS

DESCRIPTION

The program sclite is a tool for scoring and evaluating the output of speech recognition systems. Sclite is part of the NIST SCTK Scoring Tookit. The program compares the hypothesized text (HYP) output by the speech recognizer to the correct, or reference (REF) text. After comparing REF to HYP, (a process called alignment), statistics are gathered during the scoring process and a variety of reports can be produced to summarize the performance of the recognition system.

THE ALIGNMENT PROCESS

The Alignment process consists of two steps: 1) selecting matching REF and HYP texts, and 2) performing an alignment of the reference and hypothesis texts.

Step 1: Selection of matching REF and HYP texts

Sclite accepts as input a wide variety of file formats. The type of input formats define the algorithm for selecting matching REF and HYP texts. Currently sclite uses four algorithms:

Utterance ID Matching:

Input reference and hypothesis files in "trn" transcript format can be aligned by either dynamic programming (DP) or GNU's "diff".

When alignments are performed via DP, corresponding REF and HYP records with the same utterance id's are located in the REF and HYP files. DP Alignment and scoring are then performed on each pair of records. Only the utterance ID's present in the HYP file are aligned and scored. This means the REF file may contain more utterance records than the HYP.

When "diff" is used for alignment, corresponding REF and HYP records with the same utterance id's are located in the REF and HYP files. Rather than execute "diff" for each pair of records, all matching REF and HYP pairs are re-formatted to be newline separated words and written to a temporary files. Using the two temporary files, "diff" is then called to perform a global alignment. The output of "diff" is re-chunking into REF/HYP records by applying the rule: include all words in the output stream up to and including the last word in the reference record.

The reference file can contain extra transcripts, only needed transcripts are loaded.

Word Time Mark Matching:

When both the REF and HYP files are in the "ctm" format, The first step in the alignment process is to segment both the reference and hypothesis word lists by locating common areas of silence, (i.e. the absence of a word time mark). Once completed, the resulting "segments" are aligned via dynamic programming and scored as usual.

By default, the DP alignment is performed using word-to-word distances measures of: 0, 3, 3, 4 for correct, insertions, deletions and substitutions respectively.

Optionally, the command line flag '-T' forces the alignments to be performed using time-mediated alignments.

Reference Segment Time Mark to Hypothesis Word Time Mark

When the reference file format is "stm" and the hypothesis file format is "ctm", sclite chops up the hypothesis file into regions matching the reference segments. Currently, there a two methods of chopping the hypothesis file. The method is dependent on the text alignment algorithm.

When DP alignments are performed, the hypothesis file is segmented to match the reference segments by selecting the string of hypothesized words whose times occur before the end of each reference segment. The midpoint time of a word is used to determine if the word falls within a segment. DP alignments are then performed on the selected hypothesis words and the reference segment.

If the alignments are performed via "diff", pre-process the input reference and hypothesis texts, creating temporary reference and hypothesis files with one word per line. Then use GNU's "diff" program to perform a global alignment on the word lists. The output of "diff" is re-chunked into segments for scoring. Alternate reference transcripts can not be used with "diff" alignments.

Reference Segment Time Mark to Hypothesis Text file

When the reference file format "stm" and the hypothesis file format "txt" are used as inputs, the same alignment and scoring algorithm is used as describe above under the label "Reference Segment Time Mark to Hypothesis Word Time Mark" by GNU diff alignments.

Step 2: Text Alignments

Sclite can use either of two algorithms for finding alignments between reference and hypothesis word strings. The first, and most widely accepted, uses dynamic programming (DP) and the second uses GNU's "diff", a FSF (Free Software Foundation) program for comparing text files.

Dynamic Programming string alignment:

The DP string alignment algorithm performs a global minimization of a Levenshtein distance function which weights the cost of correct words, insertions, deletions and substitutions as 0, 3, 3 and 4 respectively. The computational complexity of DP is 0(NN).

When evaluating the output of speech recognition systems, the precision of generated statistics is directly correlated to the reference text accuracy. But uttered words can be coarticulated or mumbled to where they have ambiguous transcriptions, (e.i., "what are" or "what're"). In order to more accurately represent ambiguous transcriptions, and not penalize recognition systems, the ARPA community agreed upon a format for specifying alternative reference transcriptions. The convention, when used on the case above, allows the recognition system to output either transcripts, "what are" or "what're", and still be correct.

The case above handles ambiguously spoken words which are loud enough for the transcriber to think something should be recognized. For mumbled or quietly spoken words, the ARPA community agreed to neither penalize systems which correctly recognized the word, nor penalize systems which did not. To accommodate this, a NULL word, "@", can be added to an alternative reference transcript. For example, "the" is often spoken quickly with little acoustic evidence. If "the" and "@" are alternates, the recognition system will be given credit for outputting "the" but not penalized if it does not.

The presence of alternate transcriptions represents added computational complexity to the DP algorithm. Rather than align all alternate reference texts to the hypothesis text, then choose the lowest error rate alignment, this implementation of DP aligns two word networks, thus reducing the computational complexity from 2^(ref_alts + hyp_alts) * O(N_ref * N_hyp) to O((N_ref+ref_alts) * (N_hyp+hyp_alts)).

For a detailed explanation of DP alignment, see TIME WARPS, STRING EDITS, AND MACROMOLECULES: THE THEORY AND PRACTICE OF SEQUENCE COMPARISON, by Sankoff and Kruskal, ISBN 0-201-07809-0.

As noted above, DP alignment minimizes a distance function that is applied to word pairs. In addition to the "word" alignments which uses a distance function defined by static weights, the sclite DP alignment module can use two other distance functions. The first, called Time-Mediated alignment and the second called Word-Weight-Mediated alignment.

Time-Mediated Alignment

Time-Mediated alignment is a variation of DP alignment where word-to-word distances are based on the time of occurence for individual words. Time-mediated alignments are performed when the '-T' option is exercised and the input formats for both the reference and hypothesis files are in "ctm" format.

Time-mediated alignments are computed by replacing the standard word-to-word distance weights of 0, 3, 3, and 4 with measures based on beginning and ending word times. The formulas for time-mediated word-to-word distances are:

D(correct) = | T1(ref) - T1(hyp) | + | T2(ref) - T2(hyp) |
D(insertion) = T2(hyp) - T1(hyp)
D(deletion) = T2(ref) - T1(ref)
D(substitution) = | T1(ref) - T1(hyp) | + | T2(ref) - T2(hyp) | + 0.001
Distance for an Insertion or Deletion of the NULL Token '@' = 0.001

Where,

T1(x) is the beginning time mark of word x
T2(x) is the ending time mark of word x

Word-Weight-Mediated Alignment
Word-weight-mediated alignment is a variation of DP alignments where word-to-word distances are based on pre-defined word-weights. Each word has a unique weight assigned to it, via either a word-weight-list file, using the -w option, or through a language model file, using the -L option. The formulas for word-weight-mediated word-to-word distances are:
D(correct) = 0.0
D(insertion) = W(hyp)
D(deletion) = W(ref)
D(substitution) = W(hyp) + W(ref)
Distance for and Insertion or Deletion of the NULL Token '@' = 0.001
Where W(x) is the weight assigned to word 'x'.

String alignments via GNU's "diff":

While the DP algorithm has the advantage of flexibility, it is slow for aligning large chunks of text. To address the speed concerns, an alternative string alignment module, which utilizes GNU's "diff", has been added to sclite. The sclite program pre-processes the input reference and hypothesis texts, creating temporary reference and hypothesis files with one word per line. Then GNU's "diff" program is used to perform a global alignment on the word lists and the output is re-chunked into utterances or text segments for scoring.

Alignments can be performed with "diff" in about half the time taken for DP alignments on the standard 300 Utterance ARPA CSRNAB test set. However, in the opinion of the author, "diff" has the following bad effects:

1. it can not accommodate transcription alternations,

2. "diff" does not produce the same alignments as the DP alignments,

3. there is an increase measured error rates.

After reference and hypothesis texts have been aligned, scores are tallied for each speaker and each ref/hyp pair. After the tallies are made, a variety if output reports are generated by using the '-o' option. Here is a set of examples.

The categories tallied are:

Percent of correct words	=	# Correct words # Reference words	* 100
Percent of substituted words	=	# Substituted words # Reference words	* 100
Percent of inserted words	=	# Inserted words # Reference words	* 100
Percent of deleted words	=	# Deleted words # Reference words	* 100
Percent of sentence errors	=	# incorrect ref and hyp pairs # ref and hyp pairs	* 100

A variation in scoring called Weighted-Word Scoring can also be implemented by sclite. After Word-Weight-Mediated Alignment, the word weights can be tallied to produce weighted-word scores. The formulas for weighted-word scoring are very simliar to word scoring described above. The difference is rather than assume each word has the same weight, 1 in the case of word scoring, each individual word has a different weight. The word scoring formulas become:

Weighted Percent of correct words	=	Sum of W(hyp) if correct Sum of W(ref)	* 100
Weighted Percent of substituted words	=	Sum of W(hyp) + W(ref) if substituted Sum of W(ref)	* 100
Weighted Percent of inserted words	=	Sum of W(hyp) if inserted Sum of W(ref)	* 100
Weighted Percent of deleted words	=	Sum of W(ref) if deleted Sum of W(ref)	* 100

W(hyp) is the weight assigned to a hypothesis word, and W(ref) is the weight assigned to a reference word. Optionally deletable words have the default weight of 0.0.

WORD CONFIDENCE MEASURE EVALUATION

Confidence scores for each hypothesized word were requested of the LVCSR (Large Vocabulary Speech Recognition) participants beginning with the April 1996 evaluation. Each site was asked to do its analysis of these scores which were not processed by NIST. A review meeting was held at NIST in August 1996 which resulted in a decision to apply an agreed upon standard metric.

Confidence scores as they have been implemented are associated with each hypothesized word. (The issue has been raised whether for languages such as Mandarin, where character error rate is considered the primary measure of performance, the confidence ought to be associated with characters.) The confidence score p_c, associated with a word must be in the closed interval [0,1] and presumably, given the entropy related metric defined below, in the open interval (0,1). It should represent the system's best estimate of the a posterior probability that the hypothesized word is correct. (Correct here necessarily is with respect to an alignment procedure of the reference and hypothesis word strings.)

A single metric to use in the evaluartion of confidence scores was adopted at the August meeting. This is a normalized version of the cross entropy or mutual information. Specifically, the metric is defined as:



Sclite will automatically detect the presence of confidence measures when reading in a hypothesis "ctm" file. When sclite detects the confidence scores, the report genererated by the options "-o sum" has an additional column containing the Normalized Cross Entropy (NCE).

Output graphs concerning confidence estimates are generated by using the '-C' option. A variety of graphs can be created:

1. DET Curve Example
2. Binned Histogram Example
3. Word Confidence Score Histogram Example

REVISION HISTORY

See revision.txt in the main directory of the sclite source code directory package.

EXAMPLE USES OF SCLITE

The sclite scoring utility was written to be used as a standard scoring tool for the ARPA speech recognition benchmark tests. Since evaluation paradigms have changed over the past several years, file formats and scoring proceedures have changed as well. This utility supports the following speech recognition benchmark tests:
Utterance based evaluations:
Resource Management
ATIS (Airline Travel Information Systems):
Found speech evaluations:
Hub 4 - Marketplace and Broadcast News
Hub 5 - LVCSR Switchboard

BUGS/COMMENTS

Please contact Jon Fiscus at NIST with any bug reports or comments at the email address jonathan.fiscus@nist.gov or by phone, (301)-975-3182. Please include the version number of rover, and any other relevant information such as OS, compiler, etc.