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

More documents

Recommendations

Info

For English, several studies were carried out, both by linguists and speech scientists, to identify lexical clues which can be used to detect word boundaries [Shipman and Zue 19821. The clues used were basically constraints on sequences of phonemes, also known as phonotactic constraints. In one study [Lamel and Zue 19841, sequences of consonants of the form C+ (+ indicating a sequence of one or more consonants) were identified. These sequences were used to hypothesise some of the word boundaries in several texts. It was also suggested that such clues can also be used to detect word boundaries at a broadclass level. To identify the exact location of the word boundary in the consonant string (for example, a two consonant sequence C1C2 can contain a word boundary in three positions; before the sequence, within the sequence or after the sequence), it was suggested that additional knowledge such as acoustic-phonetics can be used. Though results on actual texts were not reported, the number of word boundaries that can be detected by these clues appear to be limited. A more recent and exhaustive study on the use of the phoneme sequence constraints was done by Harrington [Harrington, Johnson and Cooper 19871 in which sequences of the types CV, VC and CVC were considered. In this study, all word- internal sequences of the given type were extracted from a dictionary. Also all possible sequences that can occur across word boundaries were found by considering all possible pairings of the words. From these word boundary sequences the word-internal sequences were removed. Thus the remaining sequences were sequences which occur only across word boundaries and these sequences were used to hypothesise word boundaries. It was found that nearly 45% of the word boundaries can be detected in an English text represented in a phonemic form with incorrect hypotheses less than 4%. In a later study [Harrington, Watson and Cooper 19891, the above sequences were used to hypothesise word boundaries in strings represented using broadclasses. It
was found that at broadclass level, the applicability of the phoneme sequence constraints is limited. Resplts showed that the number of word boundaries detected were around 2% of the word boundaries in the text with false alarms as high as 22T0. Thus the utility of the ?honeme sequence constraints to hypothesise word bouiidaries in texts containing speech signal-to-symbol conversion errors appears to be limited. For Hindi, Ohala [Ohala 19831 gave a list of valid word-internal sequences of consonants and vowels(~+ and V') which were called sequential constraints. However in these, sequences of the form CV, VC and CVC were not considered. But, Ohala does mention that some sequences such as kyi do not occur in word-internal position. As yet no study is reported on the use of these clues for detecting word boundaries in Hindi. Our studies on the use of the phoneme sequence constraints in hypothesising word boundaries are reported in chapter 5. 2.3.3 Word boundary hypothesisation techniques based on prosodic knowledge Most of the initial studies on word boundary detection focussed on the use of prosody. Many of these were studies on human perception and used listeners to determine the word boundaries in what was heard. In one study [Nakatani and Schaffer 19781 it was shown that stress patterns aid listeners in the detection of word boundaries. In this study, 'reiterant speech' was used to eliminate the phonetic information from a phrase. In reiterant speech, all syllables in a phrase are replaced by nonsense syllables like 'ma'. For example, 'Mary had a little lamb' became 'Mama ma ma mama ma'. The effect of syntax was minimised by using sentences with the same structure and by restricting the number of syllables to be replaced to three containing a noun phrase of an adjective and a noun. For example, the sentence 'The remote stream was perfect for fishing' became 'The mama ma was perfect for fishing'. Thus only prosodic cues were available to the listener. Using several such sentences,.listeners
Page 1 and 2: Dr. €3. YEGNANARAYANA Professor D
Page 3 and 4: ACKNOWLEDGEMENTS I thank my researc
Page 5 and 6: -3 SIGNIFICANCE OF WORD BOUNDARIES
Page 7 and 8: 6.6 Location of word boundaries fro
Page 9 and 10: ABSTRACT This thesis addresses the
Page 11 and 12: \ known. The lexical analyser can n
Page 13 and 14: alounddhelongtableheagsnoddedenunis
Page 15 and 16: word boundaries into a text with wo
Page 17 and 18: words. The idea is to spot symbol s
Page 19 and 20: F1 position at a vowel-consonant bo
Page 21 and 22: -a2 A REVIEW OF THE STUDIES ON, WOR
Page 23 and 24: 1. When a midclass representation w
Page 25: particular interest is the performa
Page 29 and 30: in speaking rate indicate phrase bo
Page 31 and 32: y an estimator function E from the
Page 33 and 34: studies, the following issues are i
Page 35 and 36: 2. The effect of word boundary info
Page 37: * This selection can be done in man
Page 40 and 41: me : me:ra: metra: na: me:ra: na:m
Page 42 and 43: No. of alternatives 0-10 10-100 100
Page 44 and 45: 3.3.2 Results of lexical analysk wi
Page 46 and 47: had to be restricted to the above r
Page 48 and 49: Match Sentence nunber cost 1 2 3 4
Page 50 and 51: Match Sentence nuher cost 1 2 3 4 5
Page 52 and 53: 6 10 Number of 10 a1 ternate word s
Page 54 and 55: lexlcal analysis Mismatch cost Fig.
Page 56 and 57: had all word boundaries, is plotted
Page 58 and 59: Watch Sentence number cost 1 2 3 4
Page 60 and 61: Time lor lexical analysis 0 1 2 3 4
Page 62 and 63: Match Sentence tunter cost 1 2 3 4
Page 64 and 65: Time lor lexical analysis 1 0 1 2 3
Page 66 and 67: speech recognition process. On the
Page 68 and 69: proposed clues are based on the obs
Page 70 and 71: Case Markers: Pronouns : ka:, ki:,
Page 72 and 73: form. The above Hindi text was corr
Page 74 and 75: that if there are some substitution
Page 76 and 77:
These measures were defined as foll
Page 78 and 79:
Case Conjunc- Pronouns Verb Aux . A
Page 80 and 81:
- I-' X Error in input text X Error
Page 82 and 83:
% Error Fig.4.4 Results of word bou
Page 84 and 85:
% Error Fig.4.5 A comparison of the
Page 86 and 87:
0 I 1 I 0 10 20 30 40 50 60 % Error
Page 88 and 89:
60 4 0 No. of subsentences No. of s
Page 90 and 91:
LC1: A Hindi word can end either in
Page 92 and 93:
Fig.4.10 along with the results for
Page 94 and 95:
oundaries spotted) against system s
Page 96 and 97:
from other words, such as 'national
Page 98 and 99:
dictionary contained nearly 30000 w
Page 100 and 101:
Vowel sequences Consonant sequences
Page 102 and 103:
VC' were considered. It can be obse
Page 104 and 105:
percentage of incorrect hypotheses
Page 106 and 107:
% Error in input text 0 10 20 30 40
Page 108 and 109:
% Error Fig.5.1 Results of word bou
Page 110 and 111:
hypotheses produced by the lexical
Page 112 and 113:
transformed into other nonword-inte
Page 114 and 115:
WB hyp. errors 300 - 200 - ( v+> 10
Page 116 and 117:
number of incorrect hypotheses will
Page 118 and 119:
I--- Correctness 60 8o lmprovement
Page 120 and 121:
% Error in input text 10 20 30 40 5
Page 122 and 123:
will be smaller and hence S and P(y
Page 124 and 125:
word boundary is within the consona
Page 126 and 127:
clues except for clues of types CV'
Page 128 and 129:
Pauses in speech, though small in n
Page 130 and 131:
Several studies on English speech s
Page 132 and 133:
Speaker VB hypotheses at various th
Page 134 and 135:
Speaker UF hypotheses at various th
Page 136 and 137:
Speaker UF hypotheses at various th
Page 138 and 139:
Fig.6.1 The plot of pitch frequency
Page 140 and 141:
From the results, it can be seen th
Page 142 and 143:
speakers, with the incorrect hypoth
Page 144 and 145:
durational effects. One possible ex
Page 146 and 147:
From these results, it can be seen
Page 148 and 149:
shown that the prosodic features of
Page 150 and 151:
chapter 7 WORD BOUNDARY CLUES BASED
Page 152 and 153:
consonants. Thus one can consider o
Page 154 and 155:
Speaker VB hypotheses(VB:&) at vari
Page 156 and 157:
Speaker UF hypotheses(UF:UI) at var
Page 158:
Speaker UB hypotheses(UB:B) at vari
Page 161 and 162:
high speaking rates. This is also b
Page 163 and 164:
vowels preceding word boundaries, i
Page 165 and 166:
word-final vowels than word-interna
Page 167 and 168:
8.1 Introduction PERFORMANCE OF WOR
Page 169 and 170:
Match Sentence nunber, cost 1 2 3 4
Page 171 and 172:
75 Reduction in the time for lexica
Page 173 and 174:
The reduction in the lexical analys
Page 175 and 176:
Hatch Sentence nuher cost 1 2 3 4 5
Page 177 and 178:
Chapter 9 SUMMARY AND CONCLUSIONS I
Page 179 and 180:
In the last study (described in cha
Page 181 and 182:
the language and lexical clues are
Page 183 and 184:
REFERENCES Baken R.J. and Daniloff
Page 185 and 186:
Cutler A. and Butterfield S. (199lb
Page 187 and 188:
Prakash M., Ramana Rao G,V., Chandr
Page 189 and 190:
Zelenski R. and Class F. (1983), A
show all

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

Create successful ePaper yourself

Delete template?

Save as template?