Improvements in the Computer Assisted Transcription
System of Handwritten Text Images
Ver´onica Romero, Alejandro H. Toselli, Jorge Civera and Enrique Vidal
Instituto Tecnol´ogico de Inform´atica, Universidad Polit´ecnica de Valencia
Cam´ı de Vera s/n, 46022 Valencia, Spain
Abstract. To date, automatic handwriting recognition systems are far from being
perfect. Therefore, once the full recognition process of a handwritten text image
has finished, heavy human intervention is required in order to correct the results
of such systems. As an alternative, an interactive framework has been presented in
previous works. The results obtained in these works show that significant amounts
of human effort can be saved. Here a new way to interact with this interactive
system is proposed. Now, as soon as the user points to the next system error,
the system proposes a new suitable continuation. This way, many explicit user
corrections are avoided. Empirical results suggest that the new interaction method
can lead to further improvements in user productivity.
1 Introduction
Many documents used every day include handwritten text and, in many cases, it would
be interesting to recognize these text images automatically. To date, automatic hand-
writing recognition systems (HTR) have proven to be suitable for restricted applica-
tions with very limited vocabulary (reading of postal addresses or bank checks) or con-
strained handwriting (forms) achieving in these kinds of tasks relatively high recog-
nition rates. However, in the case of unconstrained transcription applications (such as
old manuscripts or spontaneous sentences), the current HTR technology typically only
achieves results which do not met the quality requeriments of practical applications.
In these cases, to obtain high quality transcriptions it is necessary a post editing
process, where a human transcriptor intervention is required to check and correct the
mistakes made by the HTR system. This post-editing solution is rather uncomfortable
and inefficient for the human corrector.
In previous works [1,2], an interactive scenario called “Computer Assisted Tran-
scription of Text Images” (CATTI) has been presented. In this scenario, the system uses
the text image and a previously validated part (prefix) of its transcription to propose
a suitable continuation of the transcription. Then the user finds and correct the next
system error, thereby providing a longer prefix which the system uses to suggest a new,
hopefully better continuation. The results obtained show that this system can save sig-
nificant amounts of human effort.
In this work, a change in the CATTI user interaction is studied. Now, as soon as the
user points to the next system error, the system proposes a new suitable continuation.
Romero V., H. Toselli A., Civera J. and Vidal E. (2008).
Improvements in the Computer Assisted Transcription System of Handwritten Text Images.
In Proceedings of the 8th International Workshop on Pattern Recognition in Information Systems, pages 103-112
Copyright
c
 SciTePress

This way, many explicit user corrections are avoided. To allow for an efficient imple-
mentation of this interaction improvement, a search strategy based on word-graphs is
adopted. This allows a simple modification of standard n-best lists decoding to take into
account the information provided by each error pointed by the user.
2 Foundations of CATTI
This section reviews the approach to CATTI presented in [1, 2]. The process starts when
the HTR system proposes a full transcription ˆs of a feature vectors sequence x, extracted
from a handwritten text line image. Then, the human transcriptor (named user from now
on) reads this transcription until he or she finds a mistake; i.e, he or she validates a prefix
p
′
of the transcription which is error-free. Now, the user can enter a word, c, to correct
the erroneous text that follows the validated prefix. This action produces a new prefix
p (the previously validated prefix, p
′
, followed by c). Then, the HTR system takes into
account the new prefix to suggest a suitable continuation to this prefix (i.e., a new ˆs),
thereby starting a new cycle. This process is repeated until a correct, full transcription t
of x is accepted by the user.
2.1 Formal Framework
The traditional handwritten text recognition problem can be formulated as the problem
of finding a most likely word sequence, ˆw, for a given handwritten sentence image
represented by a feature vector sequence x, that is:
ˆw = arg ma x
w
P r(w|x) = arg max
w
P r(x|w) · P r(w) . (1)
P r(x|w) is typically approximated by concatenated character Hidden Markov Mod-
els [3,4]) and P r(w) is usually approximated by a n-gram word language model [3].
In the CATTI framework, in addition to the given feature sequence, x, a prefix p of
the transcription is available and the HTR should try to complete this prefix by searching
for a most likely suffix ˆs as:
ˆs = arg max
s
P r(s|x, p) = arg max
s
P r(x|p, s) · P r(s|p) . (2)
Eq. (2) is very similar to (1), being w the concatenation of p and s. The main difference
is that now p is given. Therefore, the search must be performed over all possible suffixes
s of p and the language model probability P r(s|p) must account for the words that can
be written after the prefix p. Following assumptions and developments carried out in [1,
2] we can write:
ˆs ≈ arg max
s
max
1≤b≤m
P r(x
b
1
|p) · P r(x
m
b+1
|s) · P r(s|p) . (3)
This optimization problem entails finding an optimal boundary point,
ˆ
b, associated with
the optimal suffix decoding, ˆs. That is, the signal x is split into two segments, x
p
= x
ˆ
b
1
and x
s
= x
m
ˆ
b+1
and the search for the best transcription suffix that completes a prefix p

can be performed just over segments of the signal corresponding to the possible suffixes.
On the other hand, we can take advantage of the information coming from the prefix to
tune the language model constraints modelled by P r(s|p).
As discussed in [5], the simplest way to deal with P r(s|p) is to adapt an n-gram
language model to cope with the consolidated prefix. Assuming an n-gram model is
used for P r(w), leads to the following decomposition:
P r(s|p) ≃
k+n−1
Y
i=k+1
P r(w
i
|w
i−1
i−n+1
) ·
l
Y
i=k+n
P r(w
i
|w
i−1
i−n+1
) , (4)
where the consolidated prefix is w
k
1
= p and w
l
k+1
= s is a possible suffix. The first
term of Eq. (4) accounts for the probability of the n − 1 words of the suffix, whose
probability is conditioned by words from the validated prefix, and the second one is the
usual n-gram probability for the rest of the words in the suffix.
3 Searching
In previous works, a Viterbi-based approach was used to solve the search problem cor-
responding to Eq. 3 and 4. In this section, a more efficient approach is proposed.
As discussed in [5], we can explicitly rely on Eq. (3) to implement a decoding
process in one step, as in conventional HTR systems. The decoder is forced to match
the previously validated prefix p and then continue searching for a suffix ˆs according to
the constraints of Eq. (4). In the present work, more efficient search techniques based
on word-graphs are used. These techniques are similar to those described in [6, 7] for
Computer Assisted Translation and for multimodal speech post-editing.
A word graph represents the transcriptions with higher P r(w|x) of the given image
text sentence. In this case, the word graph is just (a pruned version of) the Viterbi search
trellis obtained when transcripting the whole image sentence. Fig. 2 shows an example
of a word graph. During the CATTI process the system makes use of this word graph in
order to complete the prefixes accepted by the human transcriptor.
A word graph can be represented as a weighted directed acyclic graph, where each
edge (e) is labeled with a word (w
e
) and a score (score(e)), and each node (n) is labeled
with a point (horizontal position) of the handwritten image (t
n
). For each edge, we
denote S
e
, E
e
as its start node and end node respectively. The graph has a single start
node, that points to the start of the text image, and a single end node.
The word labels of any path from the start node to the end node form a transcription
hypothesis, whose probability is as given in the Eq. (1). In practice, the simple mul-
tiplication of P r(x|w) and P r(w) is modified to balance the absolute values of both
probabilities. The most common modification is to use the language weight α and the
insertion penalty β as it is used in speech recognition [8]. So, we can write:
ˆw = arg max
w
log P r(x|w) + α log P r (w) + mβ , (5)
where m is the word length of w. The score of an edge is computed considering the
image between its start and end node points (x
t
E
e
t
S
e
) and the given word at the edge (w
e
):
score(e) = log P r(x
t
E
e
t
S
e
|w
e
) + α log P r(w
e
) + β . (6)

As the word graph is a representation of a subset of the possible transcriptions for a
source handwritten text image, it may happen that some prefixes given by the user can
not be exactly found in the word graph. To circumvent this problem some heuristics
need to be implemented. In this work, we modified the score associated to each edge
in order to cope with the differences between the words in the prefix and the words in
the path that best match the given prefix. This heuristic can be implemented as an error-
correcting parsing dynamic programming algorithm. Moreover, this algorithm takes
advantage of the incremental way in which the user prefix is generated, parsing only the
new suffix appended by the user in the last interaction. The modification of the score
of each edge is carried out by adding a weighted component that penalizes the score
taking into account the number of different characters (c
d
) between the word associated
to the edge (w
e
) and the word of the prefix that is being analized (w
p
):
score(e) = log P r(x
t
E
e
t
S
e
|w
e
) + α log P r(w
e
) + β + γ c
d
. (7)
Note that if w
e
= w
p
the number of different characters will be 0, therefore the equa-
tions (7) and (6) will become identical. In other case c
d
will be the minimum edit dis-
tance between w
e
and w
p
. Sometimes, it can be better delete the word associated to an
edge. In this case, we consider that the word associated to the edge is being substituted
for the empty word, so c
d
is the number of characters of w
e
. Finally, at times it can be
better to insert the word w
p
instead of substituting it for other one. In this case a new
edge is generated whose begin and end nodes are the same and whose score is:
score(e) = β + γ c
d
, (8)
where c
d
is the number of characters of w
p
. The parameter γ weights the penalization
due to the number of different characters. Its value has to be greater than 0 because,
otherwise, we will be encouraging paths which are more different from the given prefix.
The computational cost of this approach is much lower than use the na¨ıve Viterbi
adaptation we had used in previous works, because in the Viterbi adaptation the compu-
tational cost grows quadratically with the number of words of each sentence. Therefore,
using word-graph techniques the system is able to interact with the human transcriptor
in a time efficient way. However, a drawback of this implementation is that some accu-
racy can be lost.
4 Improvements in the CATTI Interaction Process
In CATTI applications the user is repeatedly interacting with the system. Hence, making
the interaction process easy is crucial for the success of the system. As it is shown in
the section 2, the interaction in the conventional CATTI consists in a mouse-click (or
equivalent pointer-positioning keystrokes) to validate the longest prefix which is error-
free, followed by typing a word to correct the erroneous text that follows the validated
prefix. In this section, a more effective way to interact with the system is presented.
Now, the mouse-click which the user makes to mark a mistake directly triggers the
system to propose a new suitable suffix.

x
INTER-0 p
ˆs antiguos cuidadores que en el castillo sus llamadas
m ↑
p
′
antiguos
INTER-1
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
ˆs cortesanos que en el castillo sus llamadas
c ciudadanos
p antiguos ciudadanos
INTER-2 ˆs que en el castillo sus llamadas
m ↑
p
′
antiguos ciudadanos que en
ˆs Castilla se llamaban
FINAL
c #
p ≡ t antiguos ciudadanos que en Castilla se llamaban
Fig.1. Example of CATTI operation. Starting with an initial recognized hypothesis ˆs, the user
validates its longest well-recognized prefix p
′
, making a mouse-click (m ), and the system emits
a new recognized hypothesis ˆs. As the new hypothesis does not correct the mistake the user
types the corrects word c, generating a new validated prefix p (c concatenated to p
′
). Taking into
account the new prefix the system suggests a new hypothesis ˆs starting a new cycle. Now, the user
validates the longest prefix p
′
with is error-free. The system takes into account the new prefix p
′
to propose a new suffix ˆs one more time. As the new hypothesis corrects the erroneous word
a new cycle start. This process is repeated until the final error-free transcription t is obtained.
Underlined boldface word in the final transcription is the only one which was corrected by user.
Note that in the iteration 1 it is needed a mouse-click to validate the longest prefix that is error-free
and then, to type the correct word. However, the iteration 2 only needs a mouse-click.
In fig. 1 we can see an example of the CATTI process with the new interaction
mode. As in the conventional CATTI, the process starts when the HTR system proposes
a full transcription ˆs of the input image x. Then, the user reads this prediction until a
transcription error is found (e) and makes a mouse-click (m) to position the cursor at
this point. This way, the user validates an error-free transcription prefix p
′
. Now, before
the user introduces a word to correct the erroneous one, the HTR system, taking into
account the new prefix and the wrong word that follows the validated prefix, suggests
a suitable continuation to this prefix (i.e., a new ˆs). If the new ˆs corrects the erroneous
word (e) a new cycle starts. However, if the new ˆs has an error in the same position that
the previous one, the user can enter a word, c, to correct the erroneous text e. This action
produces a new prefix p (the previously validated prefix, p
′
, followed by c). Then, the
HTR system takes into account the new prefix to suggest a new suffix and a new cycle
starts. This process is repeated until a correct transcription of x is accepted by the user.
It is worth noting that in the example shown in fig. 1, without interaction, a user
should have to correct about six errors from the original recognized hypothesis. If the
conventional CATTI is used the user only has to correct two words. However, with the
new interaction only one user-correction is necessary to get the final error-free tran-
scription. Note that the mouse-click m that the user makes to validate the prefix p
′
does
not involve extra human effort, because it is the same action that the user should make
in the conventional CATTI to position the cursor before typing the correct word.
This new kind of interaction needs not be restricted to a single pointer-positioning
mouse-click. If the system reaction to this mouse-click is not satisfactory (i.e., it does

Fig.2. Example of word-graph generated after the user validates the prefix “antiguos ciudadanos
que en”. The edge corresponding to the wrong-recognized word “el” was disabled.
not correct the error pointed to), the user may continue clicking and the system can react
to each click by displaying the next suffix (ordered by posterior probability) which does
not start with the already seen wrong words. It should be noted, however, that these
additional multiple clicks do involve extra user (hypothesis pondering) effort.
Since we have already dealt, in the section 2, with the problem of finding a suitable
suffix ˆs when the user validate a prefix p
′
and introduce a correct word c, we focus now
on the problem in which the user only makes a mouse-click. In this case the decoder
has to cope with the input image x, the validated prefix p
′
and the erroneous word that
follows the validated prefix e, in order to search for a transcription suffix ˆs:
ˆs = arg max
s
P r(s|x, p
′
, e) = arg max
s
P r(x|p
′
, s, e) · P r(s |p
′
, e) . (9)
Similar assumptions and developments followed in section 2 can be carried out to
model P r(x|p
′
, s, e). On the other hand, P r(s|p
′
, e) can be provided by a language
model constrained by the validated prefix p
′
and by the erroneous word that follows it.
4.1 Language Model and Search
P r(s|p
′
, e) can be approached by adapting an n-gram language model so as to cope
with the validated prefix p
′
and with the erroneous word that follows it e. The language
model presented in section 2 would provide a model for the probability P r(s|p
′
), but
now the first word of s is conditioned by e. Therefore, some changes are needed.
Let p
′
= w
k
1
be the validated prefix and s = w
l
k+1
be a possible suffix and consid-
ering that the wrong-recognized word e only affects the first word of the suffix w
k+1
,
P r(s|p
′
, e) can be computed as:
P r(s|p
′
, e) ≃ P r(w
k+1
|w
k
k+2−n
, e) ·
k+n−1
Y
i=k+2
P r(w
i
|w
i−1
i−n+1
) ·
l
Y
i=k+n
P r(w
i
|w
i−1
i−n+1
) . (10)
Now, taking into account that the first word of the possible suffix w
k+1
has to be
different to the erroneous word e, P r(w
k+1
|w
k
k+2−n
, e) can be formulated as follows:
P r(w
k+1
|w
k
k+2−n
, e) =
¯
δ(w
k+1
, e) · P r(w
k+1
|w
k
k+2−n
)
P
w
′
¯
δ(w
′
, e) · P r(w
′
|w
k
k+2−n
)
, (11)
where
¯
δ(i, j) is 0 when i = j and 1 otherwise.
As in the conventional CATTI, the decoder can be implemented using a word-graph.

The restrictions entailed by the modelling (11) can be easily implemented by deleting
the edge labeled with the word e after the prefix has been matched. An example is shown
in fig. 2. This example assumes the user has validated the prefix “antiguos ciudadanos
que en” and the wrong-recognized word was “el”. Hence, the new word-graph has the
edge labeled with the word “el” disabled.
5 HTR System Overview
The HTR system used here follows a classical architecture composed of three modules:
preprocessing, feature extraction and recognition (see [9]).
The following steps take place in the preprocessing module: first, the skew of each
page is corrected. Then, conventional noise reduction method is applied on the whole
document image, whose output is then fed to the text line extraction process which
divides it into separate text lines images. Finally, slant correction and size normalization
are applied on each separate line. More detailed description can be found in [9,10].
As our HTR system is based on Hidden Markov Models (HMMs), each prepro-
cessed line image is represented as a sequence of feature vectors. To do this, the feature
extraction module applies a grid to divide the text line image into N ×M squared cells.
In this work, N and M are chosen empirically. From each cell, three features are cal-
culated: normalized gray level, horizontal gray level derivative and vertical gray level
derivative. The way these three features are determined is described in [9]. Columns of
cells or frames are processed from left to right and a feature vector is constructed for
each frame by stacking the three features computed in its constituent cells. Hence, at
the end of this process, a sequence of M 3×N -dimensional feature vectors is obtained.
The characters are modeled by continuous density left-to-right HMMs with 6 states
and 64 Gaussian mixture components per state. Gaussians mixture serves as a proba-
bilistic law to model the emission of feature vectors of each HMM state. The optimum
number of HMM states and Gaussian densities per estate were tuned empirically.
Each lexical word is modelled by a stochastic finite-state automaton (SFS), which
represents all possible concatenations of individual characters to compose the word. On
the other hand, according to section 2, text line sentences are modelled using bi-grams,
with Kneser-Ney back-offsmoothing [11] and estimated directly from the training tran-
scriptions of the text line images.
6 Experimental Results
In order to test the effectiveness of the new way to interact with the CATTI system
different experiments were carried out. The corpora used, the different measures and
the obtained experimental results are explained in the following subsections.
6.1 Corpora
Two different corpora have been used in our experiments. The first one, called ODEC,
is a corpus based on a realistic application: transcriptions of handwritten answers ex-

tracted from survey forms made for a telecommunication company
1
. These answers
were written by a heterogeneous group of people, without any explicit or formal re-
striction. So, paragraphs become very variable and noisy. More information about this
corpus can be found in [12]. The relevant features of this corpus are shown in table 1.
Table 1. Basic statistics of the databases ODEC and CS.
ODEC CS
Number of: Training Test Total Lexicon Training Test Total Lexicon
Phrases 676 237 913 – 681 491 1,172 –
Words 12,287 4,084 16,371 3,308 6,432 4,479 10,911 3,408
Characters 64,666 21,533 86,199 80 36,699 25,460 62,159 78
The second corpus was compiled from the legacy handwriting document from the
nineteenth century identified as “Cristo-Salvador” (CS), which was kindly provided
by the Biblioteca Valenciana Digital (BIVALDI)
2
. This corpus is composed of 53 text
page images, written by only one writer. As it has been explained in section 5, the page
images have been preprocessed and divided into lines, resulting in a data-set of 1,172
text line images. A summary of relevant features of this partitions is shown in table 1.
The partition used here corresponds with the partition called “soft” in [2].
6.2 Assessment Measures
Different evaluation measures have been adopted. On the one hand, the quality of the
transcription without any system-user interactivity is given by the well known word
error rate (WER). On the other hand, the word stroke ratio (WSR) can be defined
as the number of (word level) user interactions that are necessary to produce correct
transcriptions using the CATTI system, divided by the total number of reference words.
Finally, the word click rate (WCR) can be defined as the number of additional mouse-
clicks by word that the user has to do using the new interaction with respect to using the
conventional CATTI system, also relative to the total number of words in the correct
transcription. In the experiments presented here real user interaction is simulated by
using the given reference transcriptions of the text images. Therefore, results should be
understood as estimates of expected real user effort.
The relative difference between WER and WSR (called Effort-Reduction) gives us
an estimation of the reduction in human effort achieved by using CATTI with respect
to using a conventional HTR system followed by human postediting.
6.3 Results
Table 2 shows the results obtained with the two corpora explained previously. In the
first part of the table we can see an estimation of the reduction in human effort (E-R)
achieved by using the conventional CATTI system with respect to the classic HTR post
editing. In the second part, the results obtained with the new single-click interaction
mode (explained in section 4) are shown.
1
Data kindly provided by ODEC, S.A. (www.odec.es)
2
http://bv2.gva.es

It is important to notice that some of the results in table 2 do not correspond with
those reported in ( [1, 2]). The differences are due to variations in the feature extraction
process and on the implementation of this system. In previous works, a Viterbi-based
approach was used while in this work word graphs search is used.
Table 2. Results obtained with the corpora ODEC and CS using the conventional CATTI (top)
and the new kind of single-click interaction (bottom).
ODEC Cristo-Salvador
WER (%) 25.3 33.9
WSR (%) 22.7 32.5
Estimated E-R (%) 10.3 4.1
WSR (%) 19.8 27.8
Estimated E-R (%) 21.7 18.0
0
10
20
30
40
50
60
6 5 4 3 2 1S 0
0
0.4
0.8
1.2
1.6
2
WSR|E-R
WCR
N. max. clicks
CRISTO-SALVADOR
WSR
WCR
E-R
0
10
20
30
40
50
6 5 4 3 2 1S 0
0
0.4
0.8
1.2
1.6
2
WSR|E-R
WCR
N. max. clicks
ODEC
WSR
WCR
E-R
Fig.3. WSR, Effort-Reduction (E-R) and WCR as a function of the maximal number of mouse-
clicks allowed by the user before writing the correct word. The first point (0) correspond to
the conventional CATTI, and the point S correspond to the single-click interaction discused in
section 4.
According to table 2, the estimated human effort to produce error-free transcription
using the new kind of interaction is significantly reduced with respect to using a conven-
tional HTR system or the conventional CATTI. In the ODEC task, the new interaction
mode can save about 22% of the overall effort, whereas the conventional CATTI would
only save 10.3%. In the CS corpus, the reduction achieved is about 18%, instead of 4%
obtained with the conventional CATTI.
Fig. 3 shows the WSR, the Effort-Reduction (E-R) and the WCR as a function of the
maximal number of mouse-clicks allowed by the user before writing the correct word.
The first point (0) corresponds to the results of the conventional CATTI, and the point
“S” corresponds to the the single-click interaction considered in the previous table. A
good trade-off is obtained when the maximum number of clicks is around 3, because a
significant amount of expected human effort is saved with a fairly low number of extra
clicks per word.
7 Remarks and Conclusions
In this paper,we haveproposed a new way to interact with the CATTI systems presented
in previous works. In conventional CATTI the user finds and corrects a first error and

thereby validates an error-free transcription prefix which is used by the system to pro-
pose a hopefully better transcription continuation. Now, the mouse-click with which the
user implicitly indicates the point where an error has occurred is used by the system to
attempt to correct the error pointed to. It is worth noting that alternative (n-best) suffixes
could also be obtained with the conventional CATTI system. However, by considering
the rejected words to propose the alternative suffixes, the interaction methods here stud-
ied are more effective and (hopefully) more comfortable for the user. Moreover, using
the new single-click interaction method, a second alternative suffix is obtained with-
out extra human effort. A simple implementation of this system using word-graphs has
been described and some experiments have been carried out.
In spite of the extreme difficulty of the corpora used in the experiments, the ob-
tained results suggest that this new kind of interaction can speed-up, facilitate and save
significant amounts of human effort in the handwritten text transcription process.
Acknowledgements
Work supported by the EC (FEDER), the Spanish MEC under the MIPRCV “Con-
solider Ingenio 2010” research programme (CSD2007-00018), the iTransDoc research
project (TIN2006-15694-CO2-01), and by the Universitat Polit`ecnica de Val`encia (FPI
fellowship 2006-04)
References
1. A. H., V. Romero, L.R., Vidal, E.: Computer assisted transcription of handwritten text. In:
Proc. of ICDAR 2007, Paran´a (Brazil), (IEEE Computer Society) 944–948
2. V. Romero, A. H., L.R., Vidal, E.: Computer assisted transcription for ancient text images.
In Proc. of ICIAR 2007 Vol. 4633 (2007) 1182–1193
3. Jelinek, F.: Statistical Methods for Speech Recognition. MIT Press (1998)
4. Rabiner, L.: A Tutorial of Hidden Markov Models and Selected Application in Speech
Recognition. Proc. IEEE 77 (1989) 257–286
5. Rodriguez, L., Casacuberta, F., Vidal, E.: Computer Assisted Speech Transcription. In: Proc.
of ibPRIA 2007. LNCS, Girona (Spain) (2007)
6. Barrachina, S., et al.: Statistical approaches to computer-assited translation. In: Computa-
tional Linguistic. (2006) I
7. Liu, P., Soong, F.: Word graph based speech recognition error correction by handwriting
input. Proc. of the 8th Int. Conf. on Multimodal interfaces (2006) 339–346
8. A. Ogawa, K.T., Itakura, F.: Balancing acoustic and linguistic probabilites. Proc. IEEE Conf.
Acoustics, Speech, and Signal Processing (1998) 181–184
9. Toselli, A.H., et al.: Integrated Handwriting Recognition and Interpretation using Finite-State
Models. Int. Journal of Pattern Recognition and Artificial Intelligence 18 (2004) 519–539
10. Romero, V., Pastor, M., Toselli, A.H., Vidal, E.: Criteria for handwritten off-line text size
normalization. In: Proc. of VIIP 06, Palma de Mallorca, Spain (2006)
11. Kneser, R., Ney, H.: Improved backing-off for n-gram language modeling. In Proc. of
ICASSP 1995 1 (1995) 181–184
12. Toselli, A.H., Juan, A., Vidal, E.: Spontaneous Handwriting Recognition and Classification.
In: Proc. of ICPR 2004. Volume 1., Cambridge, United Kingdom (2004) 433–436