Sanghamitra Mohanty

An Efficient Bilingual Optical Character Recognition (English-Oriya) System for Printed Documents

Recognition of documents containing multiscripts is really a challenging task, which needs more effort of the OCR (Optical Character Recognition) designers for improving the accuracy rate. Previously OCR was developed for documents with single scripts only mainly for English and regional languages. Old documents of not only uniscripts but also multiscripts is needed to be preserved for future use. This paper describes the character recognition process for printed documents containing English and Oriya texts. Though the languages in India are different but still we can find some common features among them. In consideration to our paper we need to distinguish between the Roman Script and the Oriya Script. Most of the English that is. Roman Script are linear as well as circular in nature and the Oriya characters are circular in nature. So we need to separate these scripts by taking into consideration of their features paragraph wise or line wise.

Speech input-output system in Indian farming sector

In this research paper, we have developed a Speech input-output System in Indian language that can be useful in bridging the gap between lab and land. This voice enabled system will be very helpful to Indian farmers to retrieve the cultivation information like availability of seeds, fertilizers, pesticides and methods of farming. The speech input-output system is very beneficial in two angles for Indian cultivation domain, first it needs the input query of farmers in voice mode which completely avoids the end users to be literate persons and secondly the response of the system also in voice mode which is more comprehensive for every class of end-user. Speech input-output system was developed using CMU Sphinx4 for understanding the enduser query and speech synthesis was achieved by playing the prerecorded sound files for providing the response to the spoken query. We have collected the speech corpus related to Indian cultivation for this research work in Odia language which is one Indian official language. The performance evaluation of the system is measured in terms of percentage of word accuracy and word error rate as well as using mean opinion score(MOS) test. The word accuracy for known and new users is found to be 86.87 % and 75.13% respectively. The MOS test for speech output is found as 4.62 on an average.

Vowel Analysis for Identifying Expression of Emotions in Odia- Spoken-Language

Speech can be described as an act of producing voice through the use of the vocal folds and vocal apparatus to create a linguistic act for conveying intended information. Speech is a subtle and rich communication that transfers not only the linguistic information, but also information about the personality and the emotional state of the speaker. The objective of this study is to analyse the effect of vocal characteristics like duration of vowels, mean pitch, jitter and shimmer of vowels in different Oriya spoken emotional speech uttered by female & male speakers. This paper also used FFBPNN as a classifier to identify an emotion out of five emotional categories such as Normal, Santa, Karuna, Raoudra, Shringar using Jitter and Shimmer variants as feature vectors. The supervised classifier achieved 86.88% accuracy rate in vowel domain. The emotional speech corpus is recorded from several semi-professionals actors within range of 22-35 years old.

NLP techniques in enriching e-governance

E-Governance is the use of Information and Communication Technologies for improving the functioning of government organizations to enable more citizen-centric services and relationships. It involves the engagement of citizen in the decision making processes for providing more effective services and policies. In this paper, we shall be discussing the application of Natural Language Processing (NLP) in e-Governance. NLP has the strong potentials in supporting e-governance objectives particularly in the area of localization of contents and analysis of contributions by citizens in discussion forum. In this paper we present some ideas on how NLP technologies can support localization in e-governance.

Algorithm for segmenting script-dependant portion in a bilingual Optical Character Recognition system

Documents may contain multiscript and recognition of those documents is really a challenging task. Earlier OCR (Optical Character Recognition) was developed for documents containing only English or regional languages. Documents containing multiple scripts are also needed to be kept protected for later use. So it needs more effort of the OCR designers for improving the accuracy rate for multi script OCR. In this paper we describe the character recognition process for printed documents containing Roman and Odia texts. The separation of the script has been performed at line level. We discuss a detailed description on the segmentation scheme using X-Y-Cut algorithm which isolates the text image into individual Odia and Roman line. To distinguish between the Roman and the Odia script, along with line height, we have considered the features of both the scripts. Most of the Roman character is linear as well as circular in nature and Odia characters are circular occupying more width as compared to Roman characters. We emphasize on Upper and lower Matras associated with Odia and absent in English. After extracting the individual scripts from the bilingual documents line wise, we send them to their individual OCR for recognition. Thus an algorithm has been proposed for identification of Odia and Roman scripts.

S2S system for voice oriented tourism information delivery in Indian context

In this paper we address a speech-to-speech (S2S) system in Odia language using multi-pattern recognition approaches for voiced oriented tourism information delivery. S2S provides spoken interface for human machine communication in Odia language. The presented system incorporates an HMM-based continuous speech recognizer using trigram model, K-nearest neighbour (KNN) classifier and a speech synthesis system that outputs speech related to Indian tourism information in Odia language. Three different databases are employed in S2S system especially for speech recognizer that consists of 2500 tourism continuous spoken queries collected from 50 speakers, labelled database for KNN and diphone corpus for speech synthesizer. The overall accuracy of the system is measured in multanimous ways by emphasising on the architecture of the developed system. Firstly, it was measured in terms of HMM based speech recognizer output which is calculated in the terms of sentence and word accuracy rate over old users and new users. The sentence accuracy and word accuracy found as 71.22% and 92.35 % respectively for old users where as the performance was found as 67.41% and 88.44 % for new users respectively. Secondly, KNN was applied over the filtered text output of speech recognizer to find the solution of query of tourists. The accuracy of the KNN was found as 89.23%. Finally, the pattern selected by KNN was sent to unicode based Odia speech synthesizer to produce speech output in Odia language. The speech synthesizer performance was evaluated in terms of MOS (Mean Opinion Score) test. The average MOS value was found as 4.2.

Double ended speech enabled system in Indian travel & tourism industry

In this research paper we emphasized on development of double ended voice enabled system in order to receive the voice query and convey voiced output message related to travel and tourism domain in Indian language. The voice enable system was developed using multiple components such as automatic speech recognizer (ASR) engine, query classifier and speech synthesis engine. The speech recognition engine plays very crucial role in speech based system which we have evaluated using multiple pattern recognition algorithms namely Hidden Markov Model (HMM), Support Vector Machine (SVM), ontology based feed forward back propagation neural network (OFFBPNN), dynamic time warping (DTW). The performance of SVM AND HMM were seen superior with respect to OFFBPNN, DTW which were measured in terms of word accuracy and word error rate. The output of ASR is fed to k-nearest neighbour (KNN) query classifier and the end result of classifier is finally passed to Odia speech synthesizer to deliver the response in voice mode. We have employed voice transformation technique in speech synthesis system to produce the spoken output in male, female, child and robotic voice. The developed double ended voice enabled system is operational over Odia spoken query and delivered the response in synthesized Odia voice.

Adaptive and iterative wiener filter for oriya speech processing applications

This paper addresses the performance enhancement of speech processing applications like speech recognition, speaker identification and language identification in the presence of additive noise with help of proposed adaptive and iterative Wiener filter. This paper deals with the problem of single microphone, frequency domain speech enhancement in noisy environments using Wiener filter in iterative and adaptive manner based on the speech signal statistics (mean and variance). The algorithm achieves good temporal resolution while maintaining formant and harmonic trajectories. The results of implementation of such a structure will demonstrate significant improvements in Oriya isolated word recognition, Oriya continuous digit recognition, speaker identification and language identification performance under noisy conditions. The accuracy of all above applications is increased by 3% to 8% in average due to the proposed speech enhancement technique incorporation in our ongoing research works.

UNL: a novel initiative for bridging digital divide

Enabling a user to access Internet resources in the language of his choice remains a major challenge in the field of Electronic Governance. This poster discusses the use of the United Nations sponsored Universal Networking Language (UNL) in bridging the language barrier by enabling translation of documents in one language (source language) to another language (target language). UNL stores information in a machine-independent and human language independent form. An EnConverter(EnCo) converts natural language text into an UNL document and a DeConverter (DeCo) generates natural language text from an UNL document.