word boundary- hypothesisation in hindi speech - Speech and ...

express the performance, (i) the number of alternate word sequences produced for a
given input phonemi sequence, and (ii) the time spent for matching. The study was
done in two parts. The first part is for the case when the input phoneme sequence
contained no errors. Hence only exact matching was used in the dictionary match. The
second part of the study is on the performance of the lexical match when the input was
assumed to contain errors likely in speech signal-to-symbol conversion. Hence
approximate string matching was used in the lexical match.
3.3.1 Results of lexical analysis with exact matching
The lexical analyser program was run with an input text containing 100
sentences. The sentences were of varying lengths and on average contained 12 to 13
words. The results of the lexical match were ordered as per the number of alternative
word strings and are shown in Table - 3.1.
From the results it can be observed that a large number of sentences (64 out of
100) had less than 10 alternate word sequences. Only a small fraction (3 oiit of 100)
had 1000 or more alternate word strings. The average number of alternate word strings
for a sentence was about 120. Three sentences had only a single word string matching
them. The highest number of alternate word strings matching any sentence were 2448.
These results were also used to study the effect of the length of the sentence
(both in terms of number of words and number of phonemes) on the number of word
sequences matching the sentence. Table - 3.2(a) and Table - 3.2(b) show the results of
this study. It can be observed that in general longer sentences have more alternate
word strings matching them. But length alone does not determine the number of
alternatives since sentences of same length still had widely different number of
matching word sequences. For example, in our data three 'sentences contained 24
words but the number of word sequences for them were 24,288 and 2488.