|
|
Some measures of intelligibility and comprehension
|
|
|
|
|
|
are spelled out completely by this module. For example, “19th” was pronounced
|
|
|
as “one-nine-T-H” on output. In words such as “19th” or “100-yard”, the al-
|
|
|
phabetic and numeric sections are separable and could be pronounced. However,
|
|
|
in a true alphanumeric such as “103S” or “a3c”, it is correct to spell out all of the
|
|
|
symbols.
|
|
|
|
|
|
A number of errors were also detected in the module DECOMP, which is
|
|
|
responsible for decomposing words into morphs by reference to the morph lexicon.
|
|
|
In several cases, the wrong morphs were identified, resulting in perceptible seg-
|
|
|
mental errors in the speech output. In other cases, the correct morphs were ob-
|
|
|
tained, but the stress assignment of the constituent morphs was different for the
|
|
|
morphs in isolation than for the morphs when concatenated in a polymorphemic
|
|
|
word. We also identified several words that should have been in the lexicon since
|
|
|
their pronunciation could not be handled by the existing spelling-to-sound rules.
|
|
|
|
|
|
Several errors in the operation of the spelling-to-sound rules were also
|
|
|
detected. These errors resulted in the wrong pronunciation, which was quite
|
|
|
noticeable in listening. For example, the second syllable of the word “Britain” was
|
|
|
pronounced like the second syllable in the word “maintain™.
|
|
|
|
|
|
In a number of other cases, we were able to identify problems in the operation
|
|
|
of the parser, particularly in recognizing the correct part of speech. For example,
|
|
|
the word “close” can be either an adjective or verb, each with a different pronun-
|
|
|
ciation. Several problems were also observed with the word “affect”, which can be
|
|
|
either a noun or a verb. In each of these cases, the part of speech was incorrectly
|
|
|
identified by the parser, resulting in the wrong choice in pronunciation on output.
|
|
|
|
|
|
Finally, there were several cases, especially with the Haskins anomalous sen-
|
|
|
tences, in which the parser incorrectly assigned the verb (which could also be a
|
|
|
noun) to the previous noun phrase. This error is not surprising since the parser has
|
|
|
a basic preference for noun phrases ényway, when a choice is available. However,
|
|
|
this often produced inappropriate sentence stress resulting from incorrect pitch and
|
|
|
|
|
|
segmental durations. In some cases, these differences could be readily observed,
|
|
|
whereas in others, the effects were substantially more difficult to detect even with
|
|
|
|
|
|
careful and repeated listening. These observations are consistent with an earlier
|
|
|
perceptual study of the durational rules carried out by Carlson et al. (1979). They
|
|
|
found that a deletion of a phrase boundary produced only negligible effects on
|
|
|
listeners’ evaluations of the naturalness of synthetic speech.
|
|
|
|
|
|
Some of the errors described above are considered to be relatively minor and
|
|
|
can be corrected rather easily by the simple addition of polymorphemic entries in
|
|
|
the morph lexicon. Since this evaluation was completed, a “pre-parser” has been
|
|
|
|
|
|
169
|
