Label selection Sample Clauses

Label selection. Inter-transcriber agreement was measured in terms of the number of substitutions, deletions and in- sertions required to map one transcription to another. The scoring algorithm required one sequence of labels to be the \reference" string, and computed the minimum number of substitutions, insertions and deletions that were needed to transform it to the second label sequence, the \hypothesis" string for each pair of transcriptions. We chose the reference string arbitrarily. Accuracy was computed as follows: Where ref, sub, ins, and del represent total number of reference segments, substitutions, insertions, and deletions, respectively. The average accuracy for the set of les in each language was computed using the average number of reference seg- ments, substitutions, insertions and deletions over all the les. In addition to the original analysis using the full label set, two more scores were calculated. The scores measured accuracy after mapping the labels to a less speci c set. The additional two levels were: Reduced symbol set produced by stripping diacritics but maintaining the base symbol. Broad categories representing: vowel, closure, plosive, fricative, semi-vowel, nasal, and nonspeech. Boundary placement. A non-trivial aspect to producing time-aligned transcriptions is the placement of the bound- aries between labels. The boundaries represent the location, in time, where the speech represented by the bounded label begins and ends. Some of these locations are arbitrary and must be de xxx by convention. Agreement on boundary lo- cation for segments on which the labelers agreed is reported for each language.