See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/270275814

Bangla Script: A Structural Study

Article · December 1998

CITATION READS

1 16,542

1 author:

Niladri Sekhar Dash

Indian Statistical Institute
168 PUBLICATIONS 713 CITATIONS

SEE PROFILE

All content following this page was uploaded by Niladri Sekhar Dash on 01 January 2015.

The user has requested enhancement of the downloaded file.

 Bangla Script: A Structural Study
Niladri Sekhar Dash and Bidyut Baran Chaudhuri
Computer Vision and Pattern Recognition Unit
Indian Statistical Institute, Kolkata, India
Email: [email protected]

Abstract

In this paper we have tried to analyse the shape of the graphemes used in the Bangla script (as
noted in printed documents). The study has focused on the formation of graphemes, their
structural changes in case of compound grapheme formation, contextual use of graphemes and
their allographs, statistical analyses of their occurrences in corpus and their positional and
functional roles in case of semantic changes. The purpose of this study is to understand the role
of the graphemes in the language; to show their behavioural peculiarities and to find out the
reasons of such peculiarities. Information obtained from this study may be useful for optical
character recognition, spelling checker designing, key-board designing, cryptography, language
teaching, and natural language processing in Bangla.

1. Introduction

The script is the visual representation of a natural language. It is the collection of some
unique symbols or characters known as graphemes, which are arranged in specific
patterns with appropriate punctuation marks in texts. The total set of symbols or
graphemes is known as alphabet of the language. Most of the linguistic features of a
language are retained in the script so that it can be easily read and understood. The study
of script is therefore very important for the study of any language. Moreover, this study
is useful and necessary for computer key-board design, optical character recognizer
development, language learning and teaching (both primary and secondary), speech
analysis and synthesis besides other applied and interdisciplinary studies.

Historians have identified two types of Proto-Indian scripts in India. One is the
Khorosthi script (also known as the Cuneiform script for it conical shape) which is used
in the North-Western Frontier provinces of India. The other is the Brahmi script which is
found in other provinces of the Northern India. The Brahmi script, deciphered by James
Princep in 1838, is claimed (Ganguli 1994) to be the origin of all modern Indian scripts
other than the Persian-Arabic and the Roman based ones. Historians believe that this
script was formed nearly 1000 years B.C. in India. The historical evolution of this script
has taken place differently in the Northern and the Southern India (Majumdar 1995).

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

2 Bangla Script: A Structural Study

Presently, in India there are four types of script in use as observed by experts (Majumder
1995):

(a) The scripts of Bangla, Devanagari, Gurmukhi, Assamese, etc., which are
originated from the old Brahmi script,
(b) The scripts of the Dravidian languages like Tamil, Telugu, Kannada and
Malayalam which are originated through old Battejhu.ttu and Pahllava script,
(c) The Persian-Arabic script like Urdu, Sindhi, Kashmiri etc, which are derived from
the Semitic script, and
(d) The Roman based script which is used for some tribal languages like Santali,
Mundari, etc. which have originally no script.

There exist some studies on the script of European languages, particularly for Roman
(English) (Diringer 1968). The study on modern Indian scripts is limited to the
modifications of script for ease of writing and printing. Thus Devnagari script has been
simplified to write and print compound graphemes. Some proposals on the Bangla script
modification are put forward during the past sixty years (Chatterji 1993). As a result,
some graphemes used in Bangla script have been deleted from the alphabet, and
modifications of some compound graphemes have been made. Nevertheless, the process
of script modification and simplification in Bangla lags far behind Devnagari (for Hindi)
and the compound grapheme shapes of Bangla, at present, are more complex than that of
the Devnagari.

There is a need for statistical studies from the corpus of Indian language texts in this
regard but very little work has been done so far. Years ago, the grapheme occurrence
frequency in Hindi has been calculated on a small corpus (Tripathi 1971) but occurrence
frequencies are not taken in a true graphemic way. In the Bangla script some statistical
studies on grapheme occurrences are conducted on some Bangla texts a few decades ago
(Bhattacharya 1965). We are not aware of any phonemic statistical analysis of Indian
language speech corpus except the one done by a small group in ISI, Kolkata (Chaudhuri
and Pal 1995).

Recently, however, a reasonably large corpus in Indian languages is available due to

sponsorship of the Department of Electronics (DOE), Govt. of India. As a result, more
complete, elaborate and robust statistical analyses are possible to carry out on this
corpus. For the study of the Bangla script we have taken the TDIL corpus that contains
more than 30 lakh words as well as a corpus consisting more than 5 lakh words which is
developed at the Computer Vision and Pattern Recognition Unit of Indian Statistical
Institute, Calcutta. Both the corpora are developed by collecting printed language data

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 3

from almost all disciplines of human knowledge published between the year 1981 and
1995. The corpora, under study, give a clear view about the characters used in the Bangla
script with adequate information about their shape, size, occurrence, articulation and
graphemic changes for various statistical analyses and observations.

This paper is organized in the following way: in Section 2, we highlight the basic
features of the Bangla script, in this section also we present a comparison of the Bangla
script with other Indian scripts. In Section 3, various statistical analyses on the Bangla
script is reported. In Section 4, we analyse the shape of the graphemes in isolation as
well as within words. In this section we also give tier-division of graphemes and
allographs along with some description on compound graphemes. In Section 5, we
evaluate the impact of characters on utterance in Bangla speech. The Section 6 is the
conclusion where the importance of this type of character analysis is discussed.

2. Features of the Bangla Script

The Bangla script is evolved through hand written documents (Sen 1992) and as a result
it has made some modifications on the shapes of the graphemes. But when it is first
mechanically designed and stratified for the purpose of printing, the scope of structural
modification is almost stopped. The first grapheme design for printing Bangla language
was done by Charles Wilkinson and his script was used to print A Grammar of Bengali
Language written by Nathaenial Brassi Halhead at the Hoogli district in 1778 (Banerjee
1981). Later, Panchanan Karmakar took this design from Wilkinson and modified the
script to give the present shape.

Similar to other Indian scripts, the Bangla script has also no indigenous origin. It is
highly influenced by its contemporary sister language scripts. Though it is directly
evolved from Brahmi, the influence and interpolation of other scripts of India can not be
ruled out. Structurally, the Oriya and all the Dravidian scripts are round, semi-round and
twisted, whereas other Indian scripts are conic and triangular in shape and form. The
Bangla script has both kinds of shape and structure though the similarity with the Aryan
script is more than that of the Dravidian scripts. The followings are the main features of
Bangla script as noted in our database.

(a) The Bangla graphemes are read and written from left to right direction both at
word and sentence levels.
(b) There are nine (9) vowel graphemes, two (2) diphthong graphemes, twenty (20)
vowel allographs, and thirty nine (39) consonant graphemes along with nearly 380
unique consonant grapheme clusters.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

4 Bangla Script: A Structural Study

(c) Except the vowel grapheme a (a) all other vowel graphemes have at least one
allograph. For example, the vowel grapheme  (A) has the allograph ◌, i (i) has
◌,  (I) has ◌, u (u) has ◌, (U) has ◌, (r) has ◌, e (e) has ◌ and ◌,  (ai)
has ◌, o (o) has ◌ and  (au) has ◌.
(d) In printed form the vowel grapheme e (e) has two allographs. One is without the
matra (headline) and the other is with matra (headline). The contexts of their use
are also different.
(e) Allographs can be used with consonant graphemes and consonant clusters but only
one at a time. Allographs can never be used with a vowel grapheme or another
vowel allograph. A single consonant grapheme or a cluster can use only one
allograph with it at a time.
(f) Vowel graphemes and allographs are always articulated in words. It never happens
that a vowel grapheme or an allograph is not articulated in spite of being present
with a consonant or cluster in a word.
(g) On the other hand, a consonant grapheme may be silent in articulation in certain
contexts despite being physically present within a word.
(h) Generally, the shape of an allograph is grapheme independent. However, there are
some exceptions where the shape of an allograph is changed based on the shape of
a consonant grapheme. For example, the original shape of the allograph ◌ (u) is
changed when it is used with consonant graphemes like  (g) and  (sh) and takes
shapes like g (gu) and  (shu), respectively.
(i) Consonant graphemes or clusters have no allograph. But the consonant grapheme
 (r) has two modifiers: Ñ (reph) and Ê (ra-phalA). Similarly, consonant grapheme
 (y) has É (yaphalA). These modifiers are normally used in cluster formation with
other consonant graphemes.
(j) A single grapheme most often represents a single abstract sound. But the reverse
one is not true. That means two or three graphemes may represent a single sound.
For example, both long  (I) and short i (i) represent /i/. Similarly, both long
(U) and short u (u) represent /u/. Among consonant graphemes, palatal  (sh) and
retroflex  (S) represent /ʃ/ while retroflex  (N) and dental  (n) represent /n/.
(k) Only consonant graphemes can form yuktabyAñjan varNa (consonant cluster).
Here consonant graphemes are physically joined in the operation. Clusters of three
or four consonant graphemes are also possible in Bangla. There are nearly 380
unique consonant grapheme clusters of which the cluster of two consonant
graphemes counts nearly 290, of three consonants count nearly 80 and that of four
consonants count around 10.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 5

(l) The sentence terminal marker in Bangla is (pUrNacched) "full stop". It is perhaps
identical for all the Aryan scripts. All Dravidian scripts, however, use a dot (.) like
the Roman script for the same function.
(m) Other punctuation marks used in Bangla script as well as in other Indian scripts
are directly borrowed from the Roman script through English.

The position of a vowel allograph with respect to a consonant grapheme or a consonant

cluster is not always uniform. The allographs like ◌ ,◌ and ◌ are put on the left hand
side of a consonant grapheme or a cluster; the allograph ◌ is used on left hand side and
above of a consonant grapheme or a cluster; the allograph ◌ is put on the right-hand
side of a consonant grapheme or a cluster, the allograph ◌ is used on the right hand side
and above of a consonant grapheme or a cluster; the allographs ◌, ◌ and ◌ are put at the
below of a consonant grapheme or a cluster. Even some allographs are put around the
consonant grapheme or a cluster, as in, ◌ and ◌. The list below (Table 1) shows the
forms and positions of allographs with respect to consonant graphemes.

Vowel Grapheme Vowel Allograph Use with Consonant Grapheme

a (a) Null (k)
 (A) ◌  (kA)
i (i) ◌  (ki)
 (I) ◌  (kI)
u (u) ◌ k (ku)
(U) ◌ k (kU)
(r) ◌  (kr)
e (e) ◌  (ke)
 (ai) ◌  (kai)
o (o) ◌   (ko)
 (au) ◌   (kau)
Table 1: Form and positions of vowel allographs with respect to consonants

Structurally, the Oriya and the Dravidian graphemes are round, semi-round and twisted
whereas other Indian graphemes are conic and triangular in shape and form. The list of
Bangla graphemes includes both kinds of shapes and structure though the similarity with
the Aryan graphemes is more than that of the Dravidian. Among the vowel graphemes
structurally, Bangla a (a) and  (A) are similar to those of Assamese, Devnagari,
Oriya, Tamil and Telugu; i (i) is similar to those of Assamese, Devnagari, Telugu, and

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

6 Bangla Script: A Structural Study

Kannada,  (I) is same to those of Assamese and Devnagari; u (u) and (U) are
identical to those of Assamese, Devnagari, Oriya, Gujarati and Gurmukhi; e (e) and 
(ai) are similar to those of Assamese, Oriya and Tamil; and finally, o (o) and  (au) are
same to those of Assamese, Oriya and Malayalam.

In case of vowel allographs it is interesting to note that none of the Indian scripts, both
Aryan and Dravidian, has allographic form for the vowel grapheme a (a). Other vowel
allographs in Bangla are similar to other Indian scripts in the following ways:

• ◌ is same with Assamese, Devnagari, Oriya, Gujarati, Gurmukhi and Tamil;

• ◌ is same with Assamese, Devnagari, Oriya, Gujarati and Gurmukhi;
• ◌ is similar to Assamese, Devnagari, Oriya, Gujarati, Malayalam and Gurmukhi,
• ◌ is same with Assamese, Oriya, Tamil, Kannada and Malayalam,
• ◌ is identical to Assamese, Oriya, and Malayalam; ◌ is same with Assamese,
Devnagari and Gujarati;
• ◌ is similar to Assamese, Oriya, Tamil and Malayalam;
• ◌ is same with Assamese and Oriya;
• ◌ is same with Assamese, Oriya, Tamil and Malayalam, and finally,
• ◌ is identical to Assamese, Devnagari, Oriya, Gujarati and Tamil.

The Bangla consonant graphemes are closely similar to that of Assamese with slightest
variation in case of  (r) and # (b). In case of  (r) in Bangla, the grapheme has a dot (.)
just below the lower arm whereas in Assamese the grapheme is crossed within with a
slanted line as in $. In case of # (b) in Bangla, the grapheme has no mark at its below,
whereas in Assamese there is a short slanted line parallel to the lower arm of the
grapheme as in %. The structural similarities of consonant graphemes with other Indian
scripts are as follows:
• Assamese has similarity with all Bangla graphemes except  (r) and # (b).
• Oriya has similarity with: & (ng), ' (D), ( (R), ) (Dh),  (N), * (t),  (n), + (th),,
(d), - (bh),. (l), ◌/ (~) and ◌0 (H).
• Devnagari has similarity with: (k),  (g), 1 (gh), & (ng), 2 (T), ' (D), ( (R), )
(Dh), 3 (Rh),  (N), + (th), , (d), 4 (dh),  (n), 5 (p), # (b), 6 (m), 7 (y), . (l), #
(v),  (S), 8 (s), 9 (h) and ◌0 (H).
• Gujarati has similarity with:  (g), 1 (gh), & (ng), : (ñ), 2 (T), ; (Th), ' (D),+
(th),  (n), 6 (m), 7 (y),  (S), 8 (s), ◌/ (~) and ◌0 (H).
• Gurmukhi has similarity with: 2 (T), * (t), 7 (y),  (r) and 8 (s).

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 7

• Tamil has similarity with: : (ñ),  (N), . (l) and ◌0 (H).

• Telugu has similarity with: * (t), ◌< (M) and ◌0 (H).
• Kannada has with similarity with: * (t), ◌< (M) and ◌0 (H).
• Malayalam has similarity with: (k),  (N),  (n), ◌< (M) and ◌0 (H).

One of the primary differences of the Bangla script from that of the Roman is that while
the latter has both Upper Case (i.e., capital letters) and Lower Case (i.e., small letters),
Bangla script has no such variation. This is also true to other Indian scripts. On the other
hand, Bangla and other Indian scripts have some vowel allographs as well as consonant
grapheme clusters which are not available in the Roman script. However, the Roman
script had some vowel grapheme clusters (e.g., Œ, Æ, æ, œ, etc.) which were designed
for specific purposes. These are no more in use now-a-days.

3. Statistical Analysis of Bangla Graphemes

The statistical analysis of the Bangla script is meticulously done over a corpus of two
hundred thousand words indigenously developed in our laboratory along with a list of
twenty five hundred thousand words collected from the TDIL corpus of DOE, Govt. of
India. The DOE text is compiled from the printed materials of different disciplines like
literature, social science, natural science, commerce, and mass media published in
between 1980-1990. It should be mentioned here that not all kinds of statistical analysis
are informed in this paper. For the convenience of understanding we have presented
only a few statistics along with some discussions in this section.

Char % Char % Char % Char % Char %

a 3.56  0.03 : 0.00  4.64  2.19
 4.85 9.81 2 0.63 5 8.68  0.05
i 0.76 = 0.99 ; 0.23 > 0.89 8 8.24
 0.05  2.29 ' 0.43 # 8.58 9 4.45
u 1.70 1 0.66 ) 0.12 - 1.91 ( 0.00
0.02 & 0.00  0.01 6 4.63 3 0.00
0.04 ? 1.99 * 4.50 7 3.11  0.13
e 5.43 @ 1.99 + 1.21  2.13 A 0.00
 0.18 я 2.44 , 4.47 . 1.23
o 1.52 C 0.13 4 0.72 # 0.01
Table 2: Words with a particular grapheme in first position

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

8 Bangla Script: A Structural Study

The Table 2 given above shows the number of unique words starting with a particular
grapheme in the first position as found in the Bangla corpus. The table highlights that
Bangla speakers feel comfortable to articulate words starting with velar, labial, sibilant
or vowel sounds. That is why words starting with (k) is highest in number followed
by that of 5 (p), # (b), 8 (s), e (e) and  (A) in a sequential order. All these sounds, by
nature, are easy to articulate because they take less energy or puff of air in articulation
than other sounds. It is the first premise, designed statistically, to show that Bangla is
easy to articulate and sweet to listen.

The Table 3 provides an interesting insight into the nature of the Bangla language. Out
of 20 graphemes the number of vowel is 6, semivowel 1, liquid 2, nasal 2, stop 8, and
fricative 1. The percentage of vowel includes both the percentage of their original and
allographic forms. Almost all the graphemes are soft, mellow to listen and easy to
articulate. This statistics easily establishes the common belief that the Bangla is virtually
a soft and mellowed language, easy to utter and sweet to listen, and may be easier to
learn than other languages.

Rank Grapheme %-age Rank Grapheme %-age

1  12.63 11 6 02.62
2 e 09.51 12 8 02.59
3  07.68 13 o 02.68
4 i 07.27 14 5 02.24
5  04.46 15 , 02.02
6 04.12 16  01.96
7 . 03.52 17 2 01.70
8 # 03.48 18  01.53
9 * 03.20 19 я 01.31
10 u 02.67 20  01.28
Table 3: Global percentage of grapheme occurrence in Bangla corpus

The Table 4 shows the occurrence of vowel is 41.69%, that of consonant is 51.70%, and
that of consonant cluster is 06.61% in the said Bangla corpus. It shows that both vowel
and consonant consist nearly 93.39% of the total graphemes used in the corpus. Though
we have nearly 380 grapheme clusters in the language, the use of clusters in the corpus is
quite less. If a page of a printed book contains 30,000 graphemes (having 300 words,
each word containing 10 graphemes in average) then nearly 28000 graphemes are either
vowel or consonant and the rests are clusters. It is noted that the percentage of cluster is

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 9

higher in similar contexts if the text is written in sadhu form (chaste version), which is
older than the calit form (colloquial version), which is now in regular use in Bangla. We
assume that the language is gradually becoming simplified as the consonant clusters are
being removed from the regular use of the text.

Grapheme %-age of use

Vowel graphemes 41.69 %
Consonant graphemes 51.70 %
Consonant grapheme clusters 06.61 %
Table 4: Percentage of grapheme occurrence

The Table 5 shows the percentage of allographs used in the corpus. It is found that the
use of ◌ (A) is the highest among all the allographs available in the language. A simple
query confirms that the sound /a/ is the maximally used sound among the vowel sounds
in regular Bangla speech. This observation is equally true for Hindi spoken corpus also
(Khan, Gupta and Rizvi 1991). Next comes the allographs ◌ (e), ◌ (i), ◌ (u) and ◌ (o),
respectively. The first two among these alographs are low and mid-high respectively in
cardinal vowel diagram. This proves that Bangla native speakers prefer low or mid-high
vowel sounds in speech to high or mid-low vowel sounds or diphthongs, the frequencies
of use of which come at the end in the table.

Allograph %-age of use Allograph %-age of use

◌ (A) 34.00 ◌(I) 03.64
◌ (e) 29.81 ◌ (U) 00.99
◌ (i) 18.56 ◌ (ri) 00.95
◌ (u) 06.70 ◌ (ai) 00.30
◌ (o) 04.72 ◌ (au) 00.28
Table 5: Percentage of use of vowel allographs in the Bangla corpus

The Table 6 given below shows which group of consonants are most frequently used in
the language. It shows that in Bangla the use of alveolar and dental consonant graphemes
is quite high. Next comes the labial followed by velar and nasal consonants, respectively.
The maximum use of soft and liquid consonants in the language proves the lucidity and
softness of the language.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

10 Bangla Script: A Structural Study

Consonant Group %-age of Use Consonant Group %-age of Use

Alveolar Group 24.81 Nasal Group 10.78
Dental Group 18.78 Palatal Group 07.59
Labial Group 16.83 Retroflex Group 06.23
Velar Group 12.36 Glottal Group 02.62
Table 6: Percentage of use of consonant groups in the Bangla corpus

The Table 7 shows the most frequently used consonant clusters as found in the Bangla
corpus. It shows that the cluster p (pr) is maximum in use followed by others. It is
highest in use because both the consonant grapheme 5 (p) and the consonant modifier Ê
(ra-phalA) are mostly used characters in the corpus text. In Bangla primary text books,
however, the cluster k (kS) is considered as a unique consonant grapheme. Perhaps, its
unique combination and high frequency of use have motivated the script designers to
consider it as a basic character. The same case does not happen for p (pr), because it is a
cluster which made with a kind of modifier Ê (ra-phalA) which, unlike  (S), is used
with almost all other consonant graphemes. Probably, for this reason it is not considered
as a unique grapheme.

Cluster %-age of use Cluster %-age of use

p (pr) 08.16 n (ny) 01.99
k (kS) 03.94 s (sth) 01.98
n (nt) 03.51 g (gr) 01.87
K (ngg) 02.50 #L (by) 01.73
t (tr) 02.48 j (jñ) 01.58
n (nd) 02.46 P (cch) 01.52
k (kt) 02.33 +R (rth) 01.48
s(st) 02.27 n (nn) 01.48
s (sv) 02.06 d (ddh) 01.47
W (ST) 02.01 X (shy) 01.31
Table 7: Percentage of use of consonant clusters in Bangla corpus

A question may raised in this context: why some consonant modifiers like Ê (ra-phalA),
É (ya-phalA), Ñ (reph), Å (va-phalA) etc. are specially designed in the script for cluster
formation? The counts taken from the corpus (as shown in Table 8) show that the use of
ra-phalA, ya-phalA, reph and va-phalA are very high in the corpus. The total percentage
of their use is 46.63% whereas the total percentage of use of other consonant clusters in
the corpus is 53.37%. This count supports our assumption that because of their frequent

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 11

use of in texts their unique form is designed to make the act of writing easy and simple.
There is also a possibility that these modifiers can be considered as unique consonant
grapheme in the script in near future as it happens for the cluster of k (kS).

Consonant Modifiers %-age of use

Ê (ra-phalA) 18.97
É (ya-phalA) 13.33
Ñ (reph) 11.35
Å (va-phalA) 02.98
Others 53.37
Table 8: Percentage of use consonant modifiers in the Bangla corpus

4. Structural Analysis of Graphemes

In a running Bangla text one can trace four types of grapheme:

(a) Vowel and consonant graphemes (basic characters),

(b) Vowel allographs and consonant modifiers,
(c) Compound graphemes (when a vowel allograph changes shape of a consonant
grapheme), and
(d) Consonant grapheme clusters.

In the following subsections the basic characters are dissected in isolation to find out the
unique properties by which a grapheme is different from the other. Modifications and
changes are also noted whenever these isolated graphemes are used in running texts.

4.1 Graphemes in Isolation

The structure of the basic Bangla graphemes is a mixture of straight lines, circular and
semi-circular curves, thick dots, and conic shapes - normally known as glyphs. All these
glyphs are not of equal size and length and each glyph is not used in its full length in
every occasion. Sometimes the full length of a glyph, sometimes the half of it and even
sometimes just a portion of the glyph is used for designing the basic graphemes. The
arrangement of these glyphs is not complex like that of Dravidian scripts. However, the
physical shape of some graphemes like  (I), (r), = (kh),  (g), 1 (gh), & (ng), @ (ch),
я (j), : (ñ), + (th), > (ph), - (bh),  (sh) and 8 (s) is more complex in form than other
graphemes. The reason of their complexity might be due to use of dots, curves, straight
lines, and conic glyphs in their shape formation.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

12 Bangla Script: A Structural Study

The shirorekhA (i.e., headline) is considered as an important feature as it can act as a

line of demarcation at the time of tier division of basic characters. It is the main feature
by which the basic graphemes can be grouped into two broad classes:

(i) Graphemes with shirorekhA (32 in number), and

(ii) Graphemes without shirorekhA (14 in number).

According to the arrangement of different glyphs the basic graphemes can be grouped
into three major classes:

(i) Graphemes made with linear structures arranged in different angles (15 in number)
(ii) Graphemes made with dot and curve shapes (11 in number), and
(iii) Graphemes made with both kinds of shape (26 in number).

The use of vertical line is maximum in formation of basic graphemes. Nearly 33 basic
graphemes have vertical lines in full span with them. The graphemes such as  (A),
(r) and C (jh) have used this vertical line twice - the second line is placed just parallel to
the first line. For most of the graphemes this vertical line is at the right most side of the
grapheme, as in,  (n), # (b), etc. However, there are some graphemes in the script like ?
(c), @ (ch), 2 (T), ) (Dh), 3 (Rh), etc. for which this vertical line is situated on their left
most side. Graphemes such as u (u), (U), & (ng), @ (ch), ' (D), , (d), ( (R) etc. have
half length vertical line in their shape design.

The width of a grapheme is not always proportionate to it height. For some graphemes
the width is more than height, as in,  (g), . (l),  (sh), etc. For some other graphemes
width is less than height, as in,  (N),  (n), etc. . Finally, for other graphemes the width
is nearly the same to height, as in, (k), # (b),  (r), etc.

For automatic character recognition, the information of shape analysis of each grapheme
is essential. In its first step it is noted that some graphemes are extremely similar in
shape to other graphemes, such as, a (a) is similar to  (A) except the vertical line with
the later; u (u) is similar to (U) except that extra curve with the later; o (o) is identical
with t (tt) except that shirorekhA (headline) on the later; = (kh) is almost same with +
(th) except that initial extra fold; (k) is almost identical with > (ph) except the gap on
the right-most vertical line; * (t) is similar with - (bh) without the initial fold for the
later; . (l) is same with  (n) except the an extra fold with the first one;  (sh) is same
with  (N) except that extra loop with the first one; e (e) and t (tr) are same except the

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 13

upper line with the later; # (b), ' (D), ) (Dh) and 7 (y) are same with  (r), ( (R), 3 (Rh)
and  (y), respectively, without the dot just below the later characters;  (g) is almost
same with 5 (p) except that the last one does not have the short slanted line connecting
the front end with the upper end of the vertical line; 1 (gh) is same with  (S) except
that slanted line that runs through the middle of the later.

Among the consonant grapheme clusters the forms k (kS) is nearly similar to h (hm)
except the loop which hangs on the right hand side of the right-hand vertical line of the
later character, etc. These graphemes are considered to be confusing graphemes as one
grapheme can be confused with the other in shape easily by man and machine.

The allographs of vowel graphemes, when these are used with consonant graphemes or
clusters are distributed in all three tiers. Some are distributed between upper and middle
tier, some are used only in lower tier, while some are distributed only on the middle tier
(see Sub-section 4.3). The reason for formation of these allographs may be to reduce the
recurrent use of vowel graphemes after the consonant graphemes and clusters in words.
The vowel graphemes, in comparison with their respective allographs, usually take more
time, space, and energy in writing. So the script designers, thinking that the use of an
allograph can be the best possible option for relieving a writer from the extra burden of
repetition, may have designed the allographs. It is observationally justified that the most
recurrently used allograph is most simple in shape and most suitably positioned in the
Bangla writing system.

4.2 Graphemes within Words

Sometimes some graphemes when used within words differ from their features noted in
isolation. On the contrary it can be said that their contexts can add some more features
which are not noted in some graphemes in their isolation. Moreover, these graphemes
can have some restrictions in their positional use; can have some modifications in their
original shape and size; and also can have some limitations in their functional role in the
strings, etc. For example, the vowel graphemes in their original forms are mostly used at
word-initial position. They can, however, be used at word-final position but at that
context they mostly function as emphatic markers, as in, .6i (kalamai) "the pen
itself", *6o (tumio) "you too", etc. Very rarely they are found to be used at the word-
middle position, such as, ?a. (cAalA) "tea vendor", a*e# (ataeb) "therefore", 6я
(mAIji) "mother", 6u** (mautAt) "relish", etc., and mostly in case of transliterated
foreign words, such as, яn (jAnuAri) "January", o2 (oATAr) "water", i
(Ain) "law", etc.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

14 Bangla Script: A Structural Study

Among consonant graphemes, & (ng), : (ñ), ( (R), 3 (Rh),  (y) and A (t) cannot occur
at word-initial position because in a normal situation it is quite difficult for a Bengali
speaker to articulate a word starting with any one of the consonant graphemes.

The consonant grapheme * (t) has a modifier, namely, A (t) (khaNData), which cannot
use vowel allograph. Generally, it occurs at the word-middle and word-final positions.
However, when there is a need to use a vowel allograph with this modifier, particularly
when a case marker is added to it, it changes into the original grapheme * (t) because
the modifier cannot carry the load of the vowel allograph, as in,,69A (mahat) "great" but
69\* (mahater) "of great", -#]A (bhabiSyat) "future" but -#]\* (bhabiSyater) "of
future", etc.

The consonant grapheme  (r) has two distinct graphic modifiers which occur at the
time of cluster formation. One is the Ñ (reph) which is placed in the upper tier just above
the consonant grapheme and which cannot cause any structural change of the consonant
grapheme. The other one is Ê (ra-phalA) which is placed at the lower tier just below the
consonant grapheme. In some occasions it can cause change in the original shape of the
grapheme in the middle tier as has happened for the consonant clusters like k (kr), t
(tr), _ (bhr), etc.

To understand the actual behaviour of the graphemes used in the Bangla script we need
to scrutinize their two important criteria within a word string:

(a) Tier division of characters, and

(b) Ability of compound grapheme formation.

Both the processes (i.e., tier division and compound grapheme formation) are necessary
and useful information for proper identification and recognition of each grapheme and
for identification of the methods used in compound grapheme formation.

4.3 Tier Division of Graphemes

In a running text the Bangla graphemes are arrayed in three tiers: upper tier, middle tier
and lower tier. The upper tier generally contains the signatures of the basic graphemes
and allographs along with some consonant modifiers like candrabindu and reph. The
middle tier virtually contains the bulk of the graphemes' weight and the lower tier carries
some allographs and consonant modifiers like ra-phalA and va-phalA. The graphic
representation of tier division of Bengali graphemes are given below (Fig. 1).

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 15

i  #  6 5 2
1

3
Fig.: 1: Tier-division of Bengali graphemes and allographs
(1: Upper Tier, 2: Middle Tier, 3: Lower Tier)

The above diagram clarifies the concept of tier division. Some times the in the middle
tier the vowel allographs are also accommodated. This division is required for grapheme
recognition and for structure analysis. Moreover, for automatic grapheme recognition by
computer, this tier division helps to identify a single grapheme in a string of multiple
different graphemes.

4.4 Compound Graphemes

For the convenience of our discussion we call them compound graphemes which are
formed by physically merging two or three graphemes together. They can be vowel and
consonant graphemes as well as vowel allographs and consonant modifiers. In this
process of formation some changes usually take place in the original structure of the
participating graphemes. Moreover, the change takes place only in the middle tier
mentioned above. Compared to the basic graphemes these graphemes are complex in
structure. For example, the allograph of the vowel grapheme u (u) when used with a
basic grapheme can generate three different compound shapes which are grapheme
dependent, as described below:

First, the allograph of the vowel grapheme u (u) takes a shape like ¦ while it is attached
in the right hand side of the grapheme  (r) giving a final shape like r (ru). The notable
point is that the change takes place only with this particular grapheme  (r), either in its
original shape or in its is Ê (ra-phalA) version with other consonant grapheme cluster like
dr (dru), gr (gru), r (shru), br (bru), etc. This shape of the allograph is similar to that of
allograph in Telugu and Kannada script. It has probably come into Bangla from Telugu
or Kannada as an outcome of cultural fusion between Bengal and South India.

Second, the allograph of the vowel grapheme u (u) takes a shape like ¦ while it is
attached at bottom of the consonant graphemes  (sh) and  (g) and the cluster n (nt) to
generate final shapes like  (shu), g (gu) and nt (ntu), respectively. This form of the
allograph is similar to that of Devnagari and Gujarati script.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

16 Bangla Script: A Structural Study

Last, the consonant grapheme 9 (h) takes the shape like h (hu) forcing the allograph to
merge with the grapheme. This variation is noted only with this particular consonant
grapheme which has no parallel form in any of the Indian scripts.

The allograph of long vowel grapheme (U) also go through structural change when it
is used with the consonant grapheme  (r) and in clusters with  (r) at the final position,
such as, g (gr), g (shr), etc. The allograph changes thoroughly into a shape like © and is
attached on the right hand side of the grapheme or cluster as in r (rU), gr (grU), r
(shrU), respectively. It is noted that the deformed allograph of this form is also similar
to that of Kannada and Malayalam script.

The allograph of the vowel grapheme (r) goes through directional change (not the
structural change) when used with the consonant grapheme 9 (h). Here the allograph,
changing its direction from horizontal to vertical is attached to the right hand side of the
grapheme as in h (hr).

Some consonant graphemes, when join physically with other consonant graphemes can
form consonant clusters. At the time of cluster formation the participating graphemes
may undergo three types of structural changes. Moreover, as said before, these changes
are noted only in the middle tier of the graphemes.

First, primary shape of the participating consonant graphemes are thoroughly changed
thereby forming a new compound shape, such as, k (kS), k (kt), k (kr), K (ngg), l
(ñc), t (tt), t (tr), and h (hm) (8 in number). In such case, it becomes almost impossible
to trace out the original shapes of the participating graphemes.

Second, original shapes of the participating graphemes are partly modified. This can be
either on both or on one of the participating graphemes. It is counted that for nearly 65
clusters, the shape of the first grapheme is affected where as there are nearly 90 clusters
where the shape of the last grapheme is affected. The reasons of such differences may be
that in the first occasion the phonetic property of the second grapheme of the cluster
holds the importance in articulation whereas in the second occasion the process becomes
just the reverse.

Last, for clusters of three and four graphemes (around 30 and 10, respectively) there is
virtually no change in final form for the first two graphemes. The last grapheme of the
cluster is placed either below or on the right hand side of the immediately preceding
grapheme, generally in the middle tier.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 17

In this context it should be mentioned that there are a few phonetic clusters in the
Bangla language, such as, /Tl/ /Dl/, /tl/, etc. for which there is graphemic representation
in the script. These are not discussed here as we intend to deal with the graphemic
clusters that are available in the printed script.

Some compound graphemes are modified for the purpose of transparency as well as for
easy access in typewriter and computer implementation. However, these modifications
are not universally accepted by all printing organizations using the Bangla script. So in
the corpus both old and new shapes of compound graphemes are available almost in
equal proportion. In the following Table 9 we have shown some opaque shapes and their
respective transparent shapes for compound grapheme design.

Opaque Shape Transparent Shape Opaque Shape Transparent Shape

g (gu) (gu)  (shu) (shu)
s (su) (su) h (hu) (hu)
h (hr) (hr) h (hn) (hn)
nt (ntu) (ntu) r (ru) (ru)
r (rU) (rU) s (st) (st)
Table 9: Opaque and transparent shapes of some compound graphemes in Bangla

5. Variations in Utterance

To locate different utterance variations of the graphemes within a word some utterance
rules are described by earlier scholars. Among them Rabindranath Tagore (1995), Jamil
Chaudhury (1990), Pabitra Sarkar (1992), Subhas Bhattacharya (1992), Enamul Haque
(1995), Mahbabul Haque (1995), Punya Sloka Ray (1997), Paresh Chandra Majumdar
(1998) are notable. Besides, some efforts are made for utterance regularization of
Bangla words by Calcutta University (1936/37), Paschimbanga Bangla Academy (1992),
Ananda Bazaar Patrika (1994), Bangiya Sahitya Parisad (1986) and Bangla Academy of
Bangladesh (1993). For our analysis and observation the utterance pattern around
Calcutta is considered as the standard one. In case of confusion, utterance dictionary
(Bhattacharya 1993) as well as some experts in the field is consulted.

The Bangla vowels, allographs, consonant, consonant modifiers and consonant clusters
constitute more than 300 unique graphemic forms. It has eight vowel graphemes: a (a),
 (A), i (i),  (I), u (u), (U), e (e) and o (o) to represent seven vowel sounds: /O/,
/a/, /i/, /æ/, /e/, /o/, and /u/. According to their articulations these graphemes as well as
their respective allographs can be grouped in the following way:

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

18 Bangla Script: A Structural Study

(a) One vowel grapheme denotes two vowel sounds: a (a) denotes /O/ and /o/;  (A)
(and allograph) denotes /a/ and /æ/; and e (e) (and allograph) denote /e/ and /æ/.
(b) Two vowel graphemes denote one vowel sound: i (i) and  (I) (along with their
allographs) denote /i/; u (u) and (U) (along with their allographs) denote /u/.
(c) O vowel grapheme denotes one vowel sound: o (o) (and its allograph) denotes /o/.

In major cases, articulation problem arises due to absence of an allograph for the vowel
grapheme a (a). Data shows that the vowel grapheme is used, if required, only at the
initial position of a word. It is sometimes articulated as /o/ creating confusion in its
primary utterance. Moreover, all non-allographed consonants and clusters in words can
be either articulated with /O/ or /o/ or may be simply non-vocalic. As a result, an
unaccustomed reader does not know which consonant or cluster should be articulated
with /O/ or /o/ or should be non-vocalic in utterance. In a similar manner, at word-initial
position the utterance of the vowel grapheme e (e) is either /e/ or /æ/. Moreover, there
are utterance variations of the consonant clusters. Thus, there are four major utterance
variations in Bangla script, namely:

(a) if a consonant grapheme is vocalic or not;

(b) if vocalic then whether it is uttered with /O/ or /o/;
(c) if the vowel grapheme e (e) (and its allograph) is uttered as /e/ or /æ/, and
(d) if the consonant grapheme or cluster is uttered with /O/ or /o/ sound or it is simply
non-vocalic in utterance.

In the Bangla corpus, in a rough count, there are nearly 5,000 words which are formed
without the use of any vowel allograph. These words are formed either by combining
vowels and consonants or by combining consonant graphemes (and clusters) only. It is,
therefore, difficult to determine the utterance of such words as it is difficult to determine
which character is vocalic and which character is not, and whether the character is
vocalic with /O/ or /o/ sound. Positional occurrence of the characters can determine their
actual utterance. The following observations may be made for non-allographed words:

(i) A non-allographed consonant or cluster at word-initial position is usually vocalic

with /O/ or /o/ sound.
(ii) A non-allographed consonant is mostly non-vocalic at word-final position except
those word that end in * (t).
(iii) For non-allographed words the normal vocality pattern is: v~n~v~n (where /v/
denotes vocalilty and /n/ denotes non-vocality).
(iv) The normal vocalization pattern is: /O~o~O~o~O~o/ or vice versa.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 19

It is already noted that the Bangla vowel grapheme a (a) has no allograph. Initially, the
vowel sound denoted by this grapheme was inherent with all consonant graphemes and
clusters (both in isolation and within words). Therefore, at the time og articulation, all
consonant graphemes and clusters are vocalic if not specified in other way. So it is quite
rational not to design an allograph for the vowel grapheme a (a) when other vowel
graphemes have one each. Whenever a consonant or a cluster without a vowel allograph
appears in texts, by default it is considered to be vocalic in sound. However, in present
day utterance we have many consonant graphemes and clusters within words which are
sometimes vocalic and sometimes not. Such differences create problem in determination
of vocality of consonant graphemes and clusters in words. Keeping these factors in mind
we have tried to define articulation patterns of some of the vowel graphemes, consonant
graphemes and clusters that may posit difficulties in utterance.

The utterance of  (A) and its allograph (◌) is mostly /a/ in all positions of a word.
However, at certain contexts, the allograph is uttered as /æ/ if used immediately after the
cluster j (jñ) in a word. We have found some words where the grapheme  (A) is
replaced by its allograph preceded by the semi-vowel grapheme  (y), as in, #-( (be-
ARA) > #( (beyARA) "obstinate", ,-# (do-Ab) > ,# (doyAb) "basin", #-
,5 (be-Adap) > #,5 (beyAdap) "obstinate", etc. Such replacement takes place
because the grapheme  (A) is not generally used at word-middle position in Bangla.
On the other hand, the use of  (yA) at this position is very common in Bangla.
Moreover, both are similar in pronunciation. So there is no hesitation in replacing  (A)
by  (yA). However, as exceptions, we have found a few words in the corpus where the
grapheme  (A) is used at word-medial position, as in, яn (jAnuAri) "January",
o2 (oATAr) "water", #i (beAinI) "illegal", #\k\. (beAkkele) "foolish",
etc. These are mostly transliterated foreign words.

The grapheme e (e) and its allograph has two utterance variations: /e/ and /æ/. For the
Bangla tongue it is always easier to glide from /e/ to /æ/ than from /e/ to /a/. Thus, to
relieve our tongue we replace /e/ before /a/ by /æ/ (Tagore 1995: 25). This is, however,
noted only at word-initial position. At other position, irrespective to any context, the
vowel grapheme or its allograph is always uttered as /e/. Other vowel graphemes and
their allographs show no variation in utterance, although, a negligible variation of length
(short and long) can be noted for /i/ and /u/.

The number of consonant graphemes (35) used in Bangla script is slightly more than the
number of consonant sounds (30) found in the language. That means some consonant

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

20 Bangla Script: A Structural Study

graphemes are identical in articulation. For instance, consonants like я (j) and 7 (y) are
almost similar in articulation; so do  (sh),  (S) and 8 (s);  (N) and  (n); * (t) and A
(t) etc.

Consonant graphemes are usually vocalic in isolation but when an allograph is attached
with them they usually drop their inherent vocalic properties to take up that particular
vowel sound the allograph denotes. However, the consonant graphemes 9 (h) and 3 (Rh)
(except 3 (ASARh) "rainy season") are always vocalic, while consonant grapheme A
(t) (khaNData) is always non-vocalic. Among consonant modifiers ◌/ (candrabindu) and
◌0 (bisarga) are always vocalic while ◌< (anusvAr) is non-vocalic.

The existence of yuktabyAñjanvarNa (consonant grapheme clusters) in Bangla script

can be traced back in the Brahmi script, which is considered as the mother of most of
the modern Indian scripts (Chatterji 1995: 56). The use of yuktabyAñjanvarNa is still
prevalent among most of the modern Indian language scripts including Bangla, Oriya,
Assamese, Devnagari etc. The Bangla script includes a large set of consonant clusters
which are formed by joining two or more consonant graphemes. Such physical merging
does not happen for vowel graphemes. The reason behind formation of such graphemic
clusters (eye cluster) can be traced in phonetic clusters (ear cluster), which occur quite
frequently in normal Bangla speech. In a random speech sequence, it is observed, two or
more consonant sounds (mostly at coda and onset of syllables) are combined together to
produce phonetic clusters, many of which are eventually realised as graphemic clusters.
Such graphemic clusters have now become an integrated part of the Bangla script and
writing system, and due to this reason they carry great importance in language learning,
text to speech conversion, OCR development and spelling checker system designing in
Bangla. A study on their formation and function can throw some light into the linguistic
behaviour of the native Bengali speakers.

We have found some discussions on the articulation patterns of consonant clusters in

Bangla (Bhattacharya 1993: 11-13, Sarkar 1994: 31-46, Ray 1997: 14-16) which have
given us some insights about the overall scenario of consonant cluster articulation. In
the following sections we have tried to give a more systematic study of the issues and
problems of this area. Out of total set of consonant grapheme clusters available in the
language, majority of the clusters are phonemic in nature because they follow the
general norms of utterance of the graphemes. But, there are some clusters which deviate
from the standard norms and create problems at the time of utterance. Only these
clusters are discussed here.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 21

For consonant clusters the general observation is that the cluster-final consonant is
always vocalic. A cluster-final consonant is one which occurs as the last member of a
cluster. At the time of utterance the characters generally follow the sequence of their
occurrence. However, for some clusters the sequence is slightly changed. Besides, some
modifications (e.g., deletion, addition and displacement of sound) in articulation also
occur due to contextual use of characters as discussed below:

At word-final position a non-allographed consonant cluster is always vocalic with /o/,

e.g., .p (kalanka) "defame", l (galpa) "story", ab (abda) "year", m (Arambha)
"beginning", #l (barañca) "rather", p N (prakANDa) "huge", #d (barAdda)
"allotted", #,g (bidagdha) "wise" etc. However, for some borrowed foreign words the
word-final non-allographed cluster is non-vocalic in utterance, as in, 2R (ArT) "art",
s (AgaST) "August", e.6R (elArm) "alarm", 'R (gArD) "guard", #l (bAlb)
"bulb", 8n (sans) "sons", >l (philm) "film", 9n (hind) "Hind", l (gAlph) "gulf",
9R (harn) "horn", 8R (nArs) "nurse", etc.

In case of the cluster s (sk) the sequence of the characters, due to anaptyxis, is trans-
positioned in utterance. Thus, works like #k (bAksa) "box" and k (riksA) "ricksaw"
are sometimes uttered as /basko/ and /riska/, respectively.

In case of the cluster k (kS) both the characters lose their respective individual utterance
to produce two utterance variations:

(a) At word-initial position it is uttered as /kh/, as in, k* (kSata) "wound", k

(kSAra) "alkaline", k (kSur) "hoof", k* (kSet) "field", k- (kSobh) "anger",
k* (kSati) "loss", etc.
(b) At other positions it is uttered as /kkh/, as in, ak (akSa) "orbit", k (kakSa)
"chamber", ?k (cakSu) "eye", 5k (pakSa) "wing", 5k (pakSi) "bird", 8k6
(sakSam) "able", ak6 (akSam) "unable", etc.

The cluster j (jñ) has also two utterance variations within words. In both cases the first
consonant grapheme я (j) loses its own utterance to be pronounced as /g/:

(a) At word-initial position the consonant grapheme я (j) is uttered as /g/, as in, j
(jñAn) "knowledge", j* (jñAnata) "in sense" etc.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

22 Bangla Script: A Structural Study

(b) At other positions the consonant grapheme я (j) is uttered as /gg/, as in, aj (ajña)
"idiot", #j (bijña) "wise", a-j (abhijña) "experienced" etc.

In case of cluster l (ñc), ~ (ñch) and я (ñj), the first consonant : (ñ) is uttered like
dental  (n), as in, aя (añjan) "eye-salve", l (kAñcan) "gold", #~ (bAñchA)
"wish", etc. However, in case of clusters made with reverse arrangement of consonants,
as in, € (cñ) and j (jñ), it loses its own utterance to nasalize its immediately preceding
character, as in, 7€ (yAcñA) "want", aj (ajñAn) "senseless" etc.

In clusters of  (shm) and s (sm) at word-initial position the utterance of 6 (m) is lost
to nasalize the preceding character, as in,  (shmashAn) "burning ghat", r
(shmashru) "beard", s (smar) "cupid", s (smaraN) "remember", s (smArak)
"memento", s (smer) "smiling" etc. A few exceptions are noted where 6 (m) has
retained its own utterance, as in, s* (smitA) "smiling" etc.

In case of clusters of t (tm), d (dm), s (sm) and  (shm) at word-medial and word-
final positions, the utterance of 6 (m) is lost to nasalize and double the utterance of the
preceding character, as in, t (AtmA) "soul", 5d (padma) "lotus",  (rashmi)
"rays", g (griSma) "summer", #s (bismay) "surprise" etc. There are, however, some
exceptions where the consonant 6 (m) is distinctly uttered with its preceding character,
as in, as* (asmitA) "selflessness",  (kAshmir) "Kashmir", k N (kuSmANDa)
"pumpkin" etc.

In cluster of k (kSm) the utterance of 6 (m) is totally lost, as in, .k (lakSmI)
"Laksmi", 5k (pakSma) "eye lash", 8k (sUkSma) "fine" etc. On the other hand, in case
of clusters of g (gm), k (km), l (lm), n (nm), the utterance of 6 (m) is retained mostly
unaffected, as in, 7g (yugma) "two", rk (rukminI) "a name", gl (gulma) "shrub",
яn (janma) "birth" etc.

For the clusters of h (hN) h (hn), h (hm) and h (hl), the actual orthographic sequence
of occurrence of the characters in words is just reversed in utterance. That means while
their orthographic pattern is C1C2, their articulatory pattern is C2C1, as in, a5h
(aparAhNa) "afternoon", ?h (cihna) "sign", bh (brAhma)¸ "Brahmin", яh, (jahlAd)
"executioner", etc.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 23

The labio-velar Å (va-phalA) as a cluster-final member modifies the utterance of clusters

in three ways, as shown below:

(a) At word-initial position its utterance is entirely lost, as in, j (jvar) "fever", t
(tvak) "skin", d5 (dvIp) "island", 5, (shvApad) "beast", (svapna) "dream" etc.
(b) At word-middle and word-final position, its own utterance is lost to double the
utterance of the preceding character, as in, 5k (pakva) "ripe", 8t (satva) "right",
#l (bilva) "a kind of wood apple", #8 (bishvAs) "faith", #d (bidvAn)
"learned", etc.
(c) In case of cluster of h (hv) at word-middle and word-final position, it generates
/bh/ sound, as in, h (AbhAn) "invitation", #h. (bihval) "exulted", h
(gahvar) "hole", etc. (Sarkar 1994: 43).

Due to its orthographic similarity with the labial consonant grapheme # (b), it is used at
the same places within words where the bilabial # (b) is normally used as a cluster-final
member, as in, #l (bAlb) "bulb" and #l (bilva) "a kind of wood apple", u\d (udbeg)
"anxiety" and #d (bidvAn) "wise", etc. In these cases, it is difficult to determine if the
character is to be articulated or not because while bilabial # (b) is always articulated the
labio-velar Å (va-phalA) is always silent in utterance. We need both etymological and
semantic information along with native language intuition to determine the utterance of
the character. This information is handy for developing systems for text-to-speech
conversion and language teaching.

The modifiers reph and ra-phalA of the consonant  (r) are always used with a character
within words. However, while reph occurs only at word-medial and word-final position,
ra-phalA can occur at all three positions of words. These modifiers can cause three types
of utterance variation, as noted below:

(a) A non-allographed consonant grapheme with reph at word-middle and word-final

position is a-ending or o-ending in utterance, as in, a R (arka) "sun", 6R (karma)
"worker", #R (garba) "proud", #яR (barjan) "discard", 85R (sarpa) "snake" etc.
Exception is noted in some borrowed words where the utterance of the consonant
is non-vocalic, as in, 2R (ArT) "art", 'R (gArD) "guard", 9R (harn) "horn", 8R
(nArs) "nurse", e.6R (elArm) "alarm" etc.,
(b) A non-allographed consonant with ra-phalA at word-initial position is normally
uttered with /o/ sound, as in, k6 (kramasha) "gradual", g9 (graha) "planet", p

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

24 Bangla Script: A Structural Study

(praNay) "love", b* (brata) "rites", _6 (bhram) "illusion", g6 (shram) "labour",

s# (srabaN) "secretion" etc. However, some exceptions are found where a non-
allographed consonant with ra-phalA at word-initial position is uttered with /O/
sound, as in, h, (hrad) "lake", k (kray) "buy" etc.
(c) A consonant with ra-phalA at word-medial and word-final position is doubled in
utterance, as in, a• (aghrAN) "a Bangla month", br (Abru) "cover", #k
(bakra) "crooked", 5t (pAtra) "bride", tr (shatru) "enemy", 8k (sakriya)
"active", #g9 (bigraha) "idol" etc.

The modifier ya-phalA occurs with a character at all positions within words. However,
depending on its position in a word it varies in utterance. At word-initial position it has
three utterance variations:

(a) With a consonant at word-initial position it has no utterance, as in, ?L# (cyaban)
"name", dL+R (dvyarthak) "ambiguous", dL* (dyuti) "glow", ?L* (cyuta)
"expelled", #L6 (byom) "ether" etc. There is an exception, like #Lk (byakta)
"expressed" etc.
(b) With a consonant tagged with the allograph of  (A) it is uttered as /æ/, as in,
L#. (kyAblA) "un-smart", ?L. (cyAlA) "follower", #L\6 (byAmo) "illness",
;L. (ThyAlA) "push", .Lя (lyAjA) "tail" etc.,
(c) With a non-allographed consonant it is uttered as /e/ if its following character is
tagged with /i/, as in, #Lk (bykti) "person", #L*k6 (bytikram) "exception",
#L*\ (bytirek) "difference", #L** (byAtIta) "except", #L+* (byAthita)
"hurt", etc. However, there are some exceptions, where with some non-
allographed consonants it is uttered as /æ/, as in, #L-? (bybhicAr) "lechery",
#L (byayI) "expensive" etc.

At word-middle and word-final position a non-allographed consonant with ya-phalA is

almost doubled in utterance, as in, .L (kalya) "tomorrow", s (shasya) "food grains",
5,L (padya) "poetry", 8-L* (sabhyatA) "civilization", 7L* (yogyatA) "ability", 64L6
(madhyam) "middle", 9*L (hatyA) "murder" etc. However, the doubling of sounds is of
two types as shown below:

(a) The unvoiced aspirate consonant will generate unvoiced unaspirate consonant plus
unvoiced aspirate consonant, such as, 6=L (mukhya) "main", 5;L (pAThya)

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 25

"syllabus", 5+L (pathya) "food for patient" etc. This process can be explained by
the following rule:

Orthography Utterance
C + ya-phalA → C C
[-voice] [-voice] [-voice]
[+aspirate] [-aspirate] [+aspirate]

(b) The voiced aspirate consonant will generate voiced unaspirated consonant plus
voiced aspirate consonant, as in, 4)L (dhanADDhya) "rich", #4L (bAdhya)
"forced", 8-L (sabhya) "civilized" etc. This process of change can be explained by
the following rule:

Orthography Utterance
C + ya-phalA → C C
[+voice] [+voice] [+voice]
[+aspirate] [-aspirate] [+aspirate]

(c) The modifier ya-phalA with the consonant 9 (h) at word-medial and word-final
position will generate /jjh/ sound, as in, ,h (dAhya) "inflammable", 8h (sahya)
"tolerate", #h (bAhyik) "external", gh (grAhya) "care", 6h6 (muhyamAn)
"morose" etc.

The functional roles and linguistic importance of candrabindu, anusvAr and bisarga are
primarily contextual. When these are detached from contexts they lose their independent
entity in the language.

6. Conclusion

Researchers studying the evolution of thought processes in human societies believe that
development of language and script may also influence the cognitive powers of the
members of a speech community. Since script is a form of knowledge representation,
the use of alphabets makes demands on humans to code and decode knowledge, convert
auditory sounds into visual symbols, think deductively and order words to construct
sentences ["Language Instinct": Know-how: The Telegraph: 9th Feb.,1998].

The script of a language is a form of knowledge representation. In the first few sections
of the paper we have tried to understand form and structure of some Bangla characters.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

26 Bangla Script: A Structural Study

The study is important and useful in problems related to spelling correction, speech
recognition, computational linguistics, character recognition, text preparation, language
teaching and cryptography, etc. The study of the Bangla script is not an exception. Even
from pure applied point of view this study can help primary and secondary language
learners and to know how Bangla script is designed and used in a running text.

In the last section of the paper we have tried to trace utterance peculiarities of Bangla
graphemes with an intention to use these for NLP works. It is to be noted that some
surface word forms, having identical grapheme arrangements, may be uttered differently
either for difference of meaning, or for difference of lexical category or for imitation of
utterance of foreign word forms. Moreover, a fixed sequence of grapheme arrangement
in the words does not guarantee the similar sequence of utterance. So, intimate study of
utterance of each grapheme and allograph is important for proper understanding the
utterance of words. The study is important in automatic speech recognition, speech
synthesis, etc.

Acknowledgment: We acknowledge help of Department of Electronics (DOE), Govt. of

India for providing the Bangla corpus our study presented here. Discussions with Prof.
Mrinal Kanti Nath, Dept. of Linguistics, University of Calcutta is acknowledged with
thanks.

Reference

Banerjee, Chittaranjan (1981) (Ed) Dui shataker Bangla Mudran o Prakashan (Bangla
Printing and Publication in Two centuries). Calcutta: Ananda Publishers.
Banerjee, Rakhaldas (1919) The Origin of the Bengali Script. Calcutta: Calcutta
University Press. Reprinted by Nababharat Publisher, Kolkata in 1973.
Bhattacharya, Nikhilesh (1965) Some Statistical Studies of the Bangla Language.
Unpublished Doctoral Dissertation. Calcutta: Indian Statistical Institute.
Bhattacharya, Subhas (1992) Bangla Ucchaaran Abhidhan (Bengali Pronunciation
Dictionary). Calcutta: Sahitya Sansad.
Bidyasagar, Iswar Chandra (1986) Barna Parichay (Bangla Primer). Calcutta: Sishu
Sahitya Samsad.
Chattopadhyay, Sanat Kumar (Ed.) (1986) Prasanga Bangla Bhasa (Issues on Bangla
Language). Calcutta: Paschim Banga Bangla Akademi.
Chattopadhyay, Suniti Kumar (1962) Bangala Bhasatatter Bhumika (Introduction to
Bangla Linguistics). Calcutta: Calcutta University Press.
Chattopadhyay, Suniti Kumar (1988) Bhasa Prakas Bangala Byakaran (Bangla
Grammar). Calcutta: Rupa Publications.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

 Bangla Script: A Structural Study 27

Chaudhuri, Bidyut Baran and Umapada Pal (1995) Relational Studies between Phoneme
and Grapheme statistics in modern Bangla Language. Journal of Acoustic Society
of India. Vol. 23., No. 1., Pp. 67-77.
Chaudhuri, Bidyut Baran and Umapada Pal (1996) OCR Error Detection and Correction
of an Inflectional Indian Language Script. Presented in the 13th International
Conference of Pattern Recognition, Vienna, Austria.
Chaudhuri, Bidyut Baran, Umapada Pal and Pulak Kumar Kundu (1966) Non-Word
Error Detection and Correction of an Inflectional Indian Language'. Presented in
the National Symposium on Machine Aids for Translation and Communication.
JNU, New Delhi, India.
Chaudhury, Jamil (1990) Banan o Uccharan (Letter and Pronunciation). Dhaka: Bangla
Academy Press.
Coulmas, Florian (1989) The Writing Systems of the World. Oxford: Basil Blackwell.
Dash, Niladri Sekhar (1997) Applicability of NLP in Bangla: A Linguistic Perspective.
Presented in the International CSP Workshop on Approaches to Knowledge
Representation, Jadavpur University, Calcutta, 18-20th February, 1997. (MS).
Diringer, David (1968) The Alphabet: A key to the History of Mankind. Vol. I&II.
London: Hatchinson.
Ganguli, Subrata (1995) Lipir Padanka Rekhay (On the Footsteps of Script). Calcutta:
Samatat Prakashani.
Haque, Enamul (1995) Bangla Bakdhvani : Svarup o Binyas (Bengali Speech Sounds:
Nature and Distribution). Dhaka: Ayantika.
Haque, Mahbabul (1995) Bangla Bananer Niyam (Rules of Bengali Spelling). Dhaka:
Jatiya Sahitya Prakasani.
Khan, I.; S.K. Gupta and S.H. Rizvi (1991) Statistics of Printed Hindi Text Graphemes :
Preliminary Results. Journal of IETE. Vol. 37. No. 3. Pp. 268-275.
Majumdar, Paresh Chandra (1998) Bangla Banan Bidhi (Bengali Spelling Rules).
Kolkata: Dey’s Publishing.
Majumder, Nepal (1992) (Ed) Banan Bitarka (Issues on Spelling). Calcutta: Paschim
Banga Bangla Akademi.
Majumder, Paresh Chandra (1995) Adhunik Bharatiya Bhasa Prasange (In the Context
of Modern Indian Languages). Calcutta: Dey's Publishing.
Ray, Punya Sloka (1997) Bengali Language Handbook. Calcutta: Paschim Banga
Bangla Akademi.
Sampson, Geoffrey (1985) Writing System: A Linguistic Introduction. London:
Hatchinson.
Sarkar, Pabitra (1984) Bhasa Desh Kal (Language in Space and Time). Calcutta: G.A.E.
Publishers.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998

 28 Bangla Script: A Structural Study

Sarkar, Pabitra (1992) Bangla Banan Sanskar: Samasya o Sambhabana (Bangla Spelling
Reform: Problem and Possibility). Calcutta: Chirayata Prakashan.
Sarkar, Pabitra (1993) Bangla Bhasar Yuktabyanjan. Bhasa. Vol. 1. No. 1. Pp. 23-45.
Sen, Dinanath (1993) Mudrancarca (Printing Practices). Calcutta: Paschim Banga
Bangla Akademi.
Sen, Sukumar (1993) Bhasar Itibritta (History of Language). Calcutta: Ananda
Publishers.
Tagore, Rabindranath (1995) Bangla Shabdatatta (Bangla Philology). Calcutta:
Viswabharati Prakashani.
Tripathi, J.N. (1971) A statistical analysis of Devnagari (Hindi) text graphemes. Journal
of IETE. Vo. 17. No. 1. Pp. 25-27.

Linguistics Today. Vol. 1. No. 2. Pp. 1-28. 1998.

View publication stats