You can not select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
45 lines
2.9 KiB
45 lines
2.9 KiB
Some measures of intelligibility and comprehension
|
|
|
|
13.3 Word recognition in sentences
|
|
|
|
The results of the Modified Rhyme Test using isolated words indicated very high
|
|
levels of intelligibility for the segmental output of the text-to-speech system.
|
|
However, the Modified Rhyme Test employs a closed-response set involving a
|
|
forced-choice format in what may be considered a relatively low uncertainty test-
|
|
ing situation. In the recognition and comprehension of unrestricted text, a substan-
|
|
tially broader range of alternatives is available to the listener since the response set
|
|
is open and potentially infinite in size. Moreover, the sentential context itself
|
|
provides an important contribution to intelligibility of speech, a fact that has been
|
|
known for many years (Miller et al., 1951; Miller and Isard, 1963).
|
|
|
|
To evaluate word recognition in sentence context, we decided to obtain two
|
|
quite different sets of data. One set was collected using a small number of the Har-
|
|
vard Psychoacoustic Sentences (Egan, 1948). These test sentences are all mean-
|
|
ingful and contain a wide range of different syntactic constructions. In addition,
|
|
the various segmental phonemes of English are represented in these sentences in
|
|
accordance with their frequency of occurrence in the language. Thus, the results
|
|
obtained with the Harvard sentences should provide a fairly good estimate of how
|
|
well we might expect word recognition to proceed in sentences when both seman-
|
|
tic and syntactic information is available to a listener. This situation could be con-
|
|
sidered comparable, in some sense, to normal listening conditions where “top-
|
|
down” knowledge interacts with sensory input in the recognition and comprehen-
|
|
sion of speech (see Pisoni, 1978; Marslen-Wilson and Welsh, 1978).
|
|
|
|
We also collected word recognition data with a set of syntactically normal but
|
|
semantically anomalous sentences that were developed at Haskins Laboratories by
|
|
Nye and Gaitenby (1974) for use in evaluating the intelligibility of their text-to-
|
|
speech system (see also Ingeman, 1978). These test sentences permit a somewhat
|
|
finer assessment of the availability and quality of “bottom-up” acoustic-phonetic
|
|
information and its potential contribution to word recognition. Since the materials
|
|
are all meaningless sentences, the individual words cannot be identified or
|
|
predicted from knowledge of the sentential context or semantic interpretation.
|
|
Thus, the results of these tests using the Haskins anomalous sentences should
|
|
provide an estimate of the upper bound on the contribution of strictly phonetic in-
|
|
formation to word recognition in sentence contexts. Since the response set is also
|
|
open and essentially unrestricted, we would anticipate substantially lower levels of
|
|
word recognition performance on this test than on the Harvard test; in the latter
|
|
test, syntactic and semantic context is readily available and can be used freely by
|
|
the listener at all levels of processing the speech input. In addition, the results of
|
|
|
|
157
|