Static Dictionary For Pronunciation Modeling

IJRET: International Journal of Research in Engineering and Technology
ISSN: 2319-1163
STATIC DICTIONARY FOR PRONUNCIATION MODELING

K.Subramanyam1, D. Arun Kumar2
1
Assistant Professor, IT Department, R.V.R&J.C College of Engineering, Guntur, A.P, India

Assistant Professor, ECE Department, R.V.R&J.C College of Engineering, Guntur, A.P, India
[email protected], [email protected]
Abstract
Speech Recognition is the process by which a computer can recognize spoken words. Basically it means talking to your computer and
having it correctly recognize what you are saying. This work is to improve the speech recognition accuracy for Telugu language using
pronunciation model. Speech Recognizers based upon Hidden Markov Model have achieved a high level of performance in controlled
environments. One naive approach for robust speech recognition in order to handle the mismatch between training and testing
environments, many techniques have been proposed. some methods work on acoustic model and some methods work on Language
model. Pronunciation dictionary is the most important component of the Speech Recognition system. New pronunciations for the
words will be considered for speech recognition accuracy.
----------------------------------------------------------------------***-----------------------------------------------------------------------STATEMENT OF THE PROBLEM

Most current leading edge speech recognition systems are
based on an approach called Hidden Markov Modeling
(HMM). By adding new pronunciations to the Static
Dictionary the accuracy can be improved. There are 2 methods
to add pronunciations to the static dictionary. First, calculating
Levenshtein distance between the strings in the confusion
pairs. Second, add the pronunciation by frequency of
occurrence.
Automatic Speech Recognition (ASR) or Speech-to-text
conversion is a sequential pattern recognition problem. It
comprises of three major components- acoustic models,
language model and the pronunciation dictionary, which aims
to correctly hypothesize a spoken utterance into a string of
words. During the training, the system is provided with speech
data, the corresponding transcription and a pronunciation
dictionary. At the decoding time, the acoustic models and
language models trained on the task are used along with one of
the standard dictionary (CMUdict) as lexicon. After Decoding
is completed, the confusion pairs are used as arguments for
Levenshtein Distance algorithm, which gives the maximum
number of operations required to convert one string into
another. The pair which has minimum distance will be
considered for adding to the static dictionary. Later the
decoding process will be repeated to find improved recognition
accuracy. In the other method the word which has occurred
more number of times will be considered for new entry in the
static dictionary.
Automatic base form learning:

The simple method of learning pronunciation variants is to
learn each word's various pronunciations on a word-by-word
basis. Typically a phone recognizer is used to determine
possible alternatives for each word by finding a best-fit
alignment between the phone recognition string and canonical

pronunciations provided by a Static Dictionary.
DICTIONARY REFINEMENT:
Sometimes Dictionary pruning is used to improve the Speech
Recognition accuracy. Dictionary pruning is done based on the
speech training database. we may arrive 2 types of problems
1. Words that are not included in the data do not have
information to be treated with and
2. Some words tend to keep pronunciations that were rarely
observed.
To solve the Unobserved words problem, we can use central or
summary pronunciations in the pruned Dictionary [3] . The aim
of this sort of pronunciation is to capture the phonetic contents
included in the set of pronunciation variants of each word and
to consolidate them in a reduced pronunciation set.
Enhanced Tree Clustering:

This approach is contrast to decision tree based approach,
which allows parameter sharing across phonemes [4]. In this
approach a single decision tree is grown for all sub-states. The
clustering procedure starts with all polyphones at the root.
Questions are asked regarding the identity of the center phone
and its neighboring phones. At any node, the question that
yields the highest information is chosen and the tree is split.
This process is repeated until either the tree reaches a certain
size or a minimum count threshold is crossed. Compared to the
traditional multiple-tree approach, a single tree allows more
flexible sharing of parameters any nodes can potentially be
shared by multiple phones.
__________________________________________________________________________________________
Volume: 01 Issue: 02 | Oct-2012, Available @ http://www.ijret.org
185

Present work
We tried with various approaches to improve the speech
recognition accuracy for Telugu sentences. The approaches are
Calculating distance using Levenshtein distance algorithm and
minimum distance variants are added to the Static Dictionary.
1.
2.
3.
4.
Addition of the frequently occurring errors.

Addition of variant in Language model.
Changing the probability.
Transcription Modification.
Levenshtein Distance Algorithm:

Levenshtein distance algorithm is to calculate the distance
between the variants. The variants which are having minimum
distance will be added to the Static Dictionary. Then we can
observe the improved accuracy.
Addition of the Frequently Occurring Errors:

In the result file we get confusion pairs with number of times
the error was repeated. I took frequently occurring errors and
added to the Static Dictionary. Then i got the improved
Accuracy.
ISSN: 2319-1163
In all the approaches, in which i succeeded one main

observation is that when I add new variant to the dictionary,
there is reduced error rate. Which is contradiction to other
papers [3-5]. For all approaches initially I used Praat to
eliminate the noise present in the wave files. if we have noise
in the wave files I got insertion errors, because of these errors
the error rate is increasing. After deleting noise using Praat the
error rate is decreased.
Database
The speech database consists of 24 speakers voice and each
speaker spoken 40 sentences.
The database is verified with different times with increased
training data. We observed improved accuracy.
TRAIN
ING
(numbe
r of
speaker
s)
TES ACCU ERRO

TING RACY RS (%)
(num
(%)
ber of
speak
ers)
Addition of Variant in Language Model:

I also tried to include the variant in the Language model also.
But i got reduced accuracy. So i did not tried this procedure
later.
Changing the Probability:

I tried to change probabilities of the states, because I want to
know whether the accuracy will be increased or decreased. But
i could unable to open the following files in the folders which
are in model_parameters.
(i) means.
(ii) mixture_weights.
(iii) transition_matrices .
(iv) variances.
Transcription Modification:
when I get some type of errors , I modified the Transcription
for those words, i observed improved accuracy. The following
are the Examples of errors.
MEEREKKADA- MEE
EKKADIKI EKKADA
HYDERAABAD- MEERAEMI
BHAARATHADESAM-BHAARATHEEYULANDARU
The modified Transcription words are
MEERREKKADA
EKKADAKKI
HHYDERAABAD
BHAARATTHADESAM
AFTER
DICTIONARY
ADDITION
ACCURA ERRO
CY(%) RS(%)
12
12
51.159
85.797
69.420
88.478
16
58.370
71.196
76.413
65.000
20
59.826
73.696
79.130
65.870
The following are the Results when the wave files are noisy
EXPERIMENT
WORD
NAME
ACCURACY(%)
ERRORS(%)
50
61.364
59.091
51
88.696
21.739
52
84.348
31.304
53
80.870
65.217
54
93.913
16.522
55
87.826
24.348
56
73.913
68.696
57
96.522
16.522
58
89.565
18.261
59
77.391
61.739
60
88.696
38.261
61
62.609
95.652
62
90.435
21.739
__________________________________________________________________________________________
186
ISSN: 2319-1163
After addition to the Dictionary the results are as follows
63
93.043
16.522
64
93.043
11.304
65
92.174
26.087
66
81.739
20.870
67
91.304
20.000
50
99.130
22.609
68
93.913
12.174
51
95.652
12.174
69
66.957
72.714
52
98.261
10.435
70
83.478
24.348
53
96.522
28.696
71
83.478
36.522
54
99.130
10.435
72
76.522
31.304
55
97.391
8.696
73
55.000
50.833
56
97.391
16.522
57
98.261
8.696
58
99.130
5.217
59
96.522
23.478
60
99.130
7.826
After eliminating the noise using Praat the results are as

follows
EXPER WORD
IMENT ACCURA
NAME
CY(%)
ERRORS(%)
EXPERIMENT
NAME
WORD
ACCURACY(%)
ERRORS(%)
61
94.696
30.739
50
80.870
40.000
62
98.261
6.957
51
88.696
19.130
63
98.261
6.957
52
86.957
21.739
53
81.739
45.217
64
96.522
8.696
54
91.304
18.261
65
100.000
15.652
55
92.174
14.783
66
94.783
7.826
56
87.826
21.739
67
99.130
6.957
57
94.783
12.174
68
97.391
6.957
58
93.913
10.435
69
93.913
36.522
59
84.348
37.391
70
93.913
16.522
60
96.522
10.435
71
93.913
13.913
61
73.913
53.043
72
96.522
9.565
62
92.174
13.043
73
99.130
4.348
63
93.043
12.174
64
92.174
13.043
65
94.696
22.957
66
81.739
20.870
67
93.043
14.783
68
93.913
12.174
69
78.261
51.304
70
80.870
29.565
71
85.217
21.739
72
86.957
19.130
73
93.913
9.565
CONCLUSIONS
The approaches discussed in this dissertation working well.
Initially the wave files are noisy. I removed the noise with
sound recorder .But sound recorder can remove noise present
at both the ends , not in the middle. so I used Praat to remove
the noise present in the middle part of wave files . I given these
wave files to the Sphinx tool ,i got word accuracy, errors and
confusion pairs .These confusion pairs are added to the
Dictionary using the approaches discussed in this thesis. finally
I observed improved accuracy and decreased errors.
__________________________________________________________________________________________
187

FUTURE WORK
I worked with the database of 24 speakers, each 40 sentences.
it's better to work with large database. we can try to achieve
100% accuracy for individual speakers. and also we can try for
improving accuracy for large database. Extending the work for
Dynamic Dictionary. Try to record wave files in noise less
environment.
Levenshtein Distance algorithm
Steps
1
Description
Set n to be the length of s.
Set m to be the length of t.
If n = 0, return m and exit.
If m = 0, return n and exit.
Construct a matrix containing 0..m rows and
0..n columns.
2.
Initialize
the
first
row
to
0..n.
Initialize the first column to 0..m.
Examine each character of s (i from 1 to n).
Examine each character of t (j from 1 to m).
If s[i] equals t[j], the cost is 0.
If s[i] doesn't equal t[j], the cost is 1.
Set cell d[i,j] of the matrix equal to the
minimum of:
a. The cell immediately above plus 1: d[i-1,j] +
1.
b. The cell immediately to the left plus 1: d[i,j1] + 1.
c. The cell diagonally above and to the left plus
the cost: d[i-1,j-1] + cost.
3
4.
5.
6.
7.
After the iteration steps (3, 4, 5, 6) are

complete, the distance is found in cell d[n,m].
REFERENCES:
[1] Gopala krishna Anumanchipalli, Modeling Pronunciation
Variation for Speech Recognition, M.S Thesis , IIIT
Hyderabad, February 2008.
[2] Eric John Fosler-Lussier, Dynamic Pronunciation Models
for Automatic Speech Recognition, Ph.d Thesis, University
of California, Berkeley, Berkeley, CA, 1999.
[3] Gustavo Hernandez-Abrego, Lex Olorenshaw, Dictionary
Refinement based on Phonetic Consensus and Non-uniform
Pronunciation Reduction, INTERSPEECH 2004 -ICSLP 8th
International Conference on Spoken Language Processing,
October 4-8, 2004.
[4] Hua Yu and Tanja Schultz, Enhanced Tree Clustering with
Single Pronunciation Dictionary for Conversational Speech
Recognition, in Proceedings of Eurospeech, pp. 1869-1872,
2003.
[5] Dan Jurafsky, Wayne Ward, Zhang Jianping, Keith Herold,
Yu Xiuyang, and Zhang sen, What Kind of Pronunciation
ISSN: 2319-1163
Variation is Hard for Triphones to Model, in proceedings of

ICASSP, pp. 577-580, 2001.
[6] Adda-Decker M. and Lamel L, Pronunciation Variants
Across System Configuration, Speech Communication,
1999.
[7] J. M. Kessens, M. Wester, and H. Strik , Improving the
Performance of a Dutch csr by Modeling within-word and
cross-word Pronunciation Variation, Speech Commun., vol.
29, no. 2-4, pp. 193207, 1999.
[8] R. Rosenfeld, Optimizing Lexical and N-gram Coverage
via Judicious use of
Linguistic data, in Proc.
Eurospeech, (Madrid), 1995.
[9] Sk . Akbar, Comparing Speech Recognition Accuracy
using HMM and Multi-layer Perceptrons , M.Tech
dissertation, June 2008.
[10] kenneth A. Kozar, Representing Systems with Data Flow
Diagrams, Spring, 1997.
http://spot.colorado.edu/~kozar/DFD.html
[11] Finke Michael and Waibel Alex, Speaking Mode
Dependent Pronunciation
Modeling in Large Vocabulary
Conversational Speech Recognition, in Procedures of
Eurospeech, 1995.
[12] M. Ravishankar and M. Eskenazi, Automatic Generation
of Context-Dependent Pronunciations, in Proc. Eurospeech
97, (Rhodes, Greece), pp. 24672470, 1997.
[13] T. Sloboda and A. Waibel, Dictionary Learning for
Spontaneous Speech Recognition, in Proc. ICSLP 96,
(Philadelphia), pp. 23282331, 1996.
[14] A. Xavier and D. Christian, Improved Acoustic-Phonetic
Modeling in Philip's Dictation System by Handling Liaisons
and Multiple Pronunciations , in Proc. Eurospeech 95,
(Madrid), pp. 767 770, 1995.
[15] P. S. Cohen and R. L. Mercer, The Phonological
Component of an Automatic Speech Recognition System ,
Reddy , D.R. (Ed) Speech Recognition. Invited Papers
Presented at the 1974 IEEE Symposium., pp. 275319, 1975.
[16] N. Cremelie and J.-P. Martens, In Search of Better
Pronunciation Models for Speech Recognition, Speech
Communication, vol. 29, no. 2-4, pp. 115136, 1999.
[17] B. C. S. and Y. S. J., Pseudo-Articulatory Speech
Synthesis for Recognition using Automatic Feature Extraction
from X-ray Data, in Proc. ICSLP 96, (Philadelphia), pp. 969
972, 1996.
[18] I.Amdal, F. Korkmazskiy, and A. C. Surendran, Joint
Pronunciation Modeling of Non-native Speakers using DataDriven Methods, in Proc. ICSLP 00, (Beijing, China), pp.
622625, 2000.
[19] M. Bacchiani and M. Ostendorf, Joint Lexicon, Acoustic
Unit Inventory and Model Design, Speech Commun., vol. 29,
no. 2-4, pp. 99114, 1999.
[20] T. Fukada, T. Yoshimura, and Y. Sagisaka, Automatic
Generation of Multiple Pronunciations based on Neural
Networks, Speech Commun., vol. 27, no. 1, pp. 6373, 1999.
[21] S. Greenberg, Speaking in Shorthand A Syllablecentric
Perspective
for Understanding Pronunciation
Variation, in Proc. of the ESCA Workshop on Modeling
__________________________________________________________________________________________
188
ISSN: 2319-1163
Pronunciation Variation for Automatic Speech Recognition,

(Kekrade, Netherlands, May 1998. ESCA.), 1998.
[22] T. Holter and T. Svendsen, Maximum Likelihood
Modeling of Pronunciation Variation, Speech Commun., vol.
29, no. 2-4, pp. 177191, 1999.
[23] H. Strik and C. Cucchiarini, Modeling Pronunciation
Variation for ASR: a survey of the literature, Speech
Commun., vol. 29, no. 2-4, pp. 225246, 1999.
[24] M. Riley, W. Byrne, M. Finke, S. Khudanpur, A. Ljolje, J.
McDonough, H. Nock, M. Saraclar, C. Wooters, and G.
Zavaliagkos, Stochastic Pronunciation Modeling from HandLabeled Phonetic Corpora, Speech Commun., vol. 29, no. 2-4,
pp. 209224, 1999.
__________________________________________________________________________________________
189

Static Dictionary For Pronunciation Modeling

Uploaded by

Copyright:

Available Formats

Static Dictionary For Pronunciation Modeling

Uploaded by

Document Information

Original Description:

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Static Dictionary For Pronunciation Modeling

Uploaded by

Copyright:

Available Formats

IJRET: International Journal of Research in Engineering and Technology