|
|
Some measures of intelligibility and comprehension
|
|
|
|
|
|
were presented only after all of the subjects in a group had finished responding to
|
|
|
the previous test sentence, and had indicated this to the experimenter. A short
|
|
|
break was taken halfway through a testing session, after completion of the first 50
|
|
|
|
|
|
sentences.
|
|
|
|
|
|
13.3.2 Results and discussion
|
|
|
|
|
|
The responses were scored only for correct word recognition at this time. Phonetic
|
|
|
errors, when they occurred, were not considered in the present analyses, although
|
|
|
we expect to examine these in some detail at a later time. Each subject receiving
|
|
|
the Harvard sentences provided a total of 500 responses, while each subject receiv-
|
|
|
ing the Haskins anomalous sentences provided 400 responses to the final analysis.
|
|
|
|
|
|
Performance on the Harvard sentences was quite good with an overall mean
|
|
|
of 93.2 percent correct word recognition across all 21 subjects. The scores on this
|
|
|
test ranged from a low of 80 percent to a high of 97 percent correct recognition.
|
|
|
Of the 6.7 percent errors observed, 30.3 percent were omissions of complete
|
|
|
words, while the remainder consisted of segmental errors involving substitutions,
|
|
|
deletions, and transpositions. In no case, however, did subjects respond with per-
|
|
|
missible nonwords that could occur as potential lexical items in English.
|
|
|
|
|
|
As expected, word recognition performance on the Haskins anomalous sen-
|
|
|
tences was substantially worse than the Harvard sentences, with a mean of 78.7
|
|
|
percent correct recognition averaged over all 23 subjects. The scores on this test
|
|
|
ranged from a low of 71 percent correct to a high of 85 percent correct. Of the
|
|
|
21.3 percent errors recorded, only 11 percent were omissions of complete words.
|
|
|
The difference in error patterns, particularly in terms of the number of omissions,
|
|
|
between the two types of sentence contexts suggests a substantial difference in the
|
|
|
subjects’ perceptual strategies in the two tests. It seems quite likely that subjects
|
|
|
used a much looser criterion for word recognition with the Haskins anomalous sen-
|
|
|
tences simply because the number of permissible alternatives was substantially
|
|
|
greater than those in the Harvard sentences. Moreover, the presence of one stan-
|
|
|
dard syntactic structure probably encouraged subjects to guess more often when
|
|
|
the acoustic cues to word identification were minimal. In addition, there seemed to
|
|
|
be evidence of semantically based intrusions in the recall data, suggesting that sub-
|
|
|
jects were attempting to assign an interpretation to the input signal even though
|
|
|
they knew beforehand that all of the sentences were meaningless.
|
|
|
|
|
|
As noted earlier, substantial learning effects occur with synthetic speech.
|
|
|
Even after an initial period of exposure, recognition performance continues to im-
|
|
|
prove. Comparisons of word recognition performance in the first and second half
|
|
|
of each of the tests indicated the presence of a reliable learning effect. For both the
|
|
|
|
|
|
159
|
