Bengali consonants-voice to text conversion using ma

61 downloads 0 Views 409KB Size Report
Bengali consonants-voice to text conversion using ma- chine learning tool. Shikhor Kumer Roy [1], Abinash Kanti Ghosh [2], Aysa Siddika Asa[3], Md. Palash ...
International Journal of Research in Computer Engineering and Electronics. Page # 1 VOl : 6 ISSUE : 1 (March 2017)

ISSN 2319-376X

Bengali consonants-voice to text conversion using machine learning tool Shikhor Kumer Roy [1], Abinash Kanti Ghosh [2], Aysa Siddika Asa[3], Md. Palash Uddin [4], Md. Rashedul Islam [5] , and Masud Ibn Afjal [6] Abstract— Voice command-based technology is getting more popular day by day. It is a common fact to use voice for searching, software locking, or instructing computer, everywhere. So today’s the main challenge is to make the computer to recognize the input voice. The aim of this proposed work is to develop such a tool to recognize Bengali consonant-voice and convert it into its corresponding text. The system takes voice input from users and then a feed forward back propagation algorithm is used to train the artificial neural network which maps this Bengali voice into text form. The system has successfully converted any single Bengali consonants-voice character to its equivalent text. Index Terms— ANN, Automatic Speech Recognition (ASR), Consonant voice, Feed forward backpropagation algorithm, Speech Application Program Interface (SAPI), Voice recognition.

——————————  ——————————

1 INTRODUCTION

S

PEECH is the old est natu ral m eans of inform ation exchange am ong hu m an beings. For centu ries, people have tried to d evelop m ach ines to u nd erstand speech as hu m an d o in natu ral w ay [1]. Mu ch of tod ay’s technology from m obile phones to personal com pu ters and beyond u tilizes speech recognition softw are. So attem pts have been m ad e to d evelop vocally interactive com pu ters to recognize voice or speech w hich is a field of com pu ter science that d eals w ith d esigning com pu ter system s that recognize spoken w ord s. It allow s a com pu ter to id entify the w ord s that a person speaks into a m icrophone or telephone [11]. The first attem pts to d evelop Au tom atic Speech Recognition (ASR) techniqu es d u ring 1950s, w hich w ere based on the d irect conversion of speech signal into a sequ ence of phonem e-like u nits bu t failed . In the 1970s, the first positive resu lts of spoken w ord recognition cam e into existence w hen general p attern m atching techniqu es w ere introd u ced . Bu t as the extension of lim ited ap plications, the statistical ap proach to ASR started to be investigated , at the sam e period [2]. Tod ay, m ost of the com pu ter based resou rces of voice reco g————————————————

[1]Shikhor Kumer Roy ([email protected]) has completed B.Sc. in Computer Science and Engineering (CSE) from Hajee M ohammad Danesh Science and Technology University (HSTU), Dinajpur, Bangladesh in 2015 (Held in 2016). [2]A binash Kanti Ghosh ([email protected]) has completed B.Sc. C.S.E. from HSTU, Dinajpur, Bangladesh in 2015 (Held in 2016). [3] Aysa Siddika A sa ([email protected]) has completed B.Sc. C.S.E. from HSTU, Dinajpur, Bangladesh in 2015 (Held in 2016). [4] M d. Palash Uddin ([email protected]) is an assistant professor in dept. of CSE in HSTU, Dinajpur, Bangladesh. [5] M d. Rashedul Islam ([email protected]) is a lecturer in dept. of CSE in HSTU, Dinajpur, Bangladesh. [6] M asud Ibn A fjal ([email protected])is an assistant professor in dept. of CSE in HSTU, Dinajpur, Bangladesh.

nition are in English. Bengali is largely spoken by the peop le all over the w orld abou t 210 m illion people speak Bengali as their native langu age [3]. It is ranked seventh based on the nu m ber of speakers and the native langu age in Banglad esh. Early researchers have d eveloped Bengali speech recognition system , bu t a very few implem entation based w orks have been d one [14]. The m ain goal of this w ork is to convert consonants-voice into its correspon d ing text for the interaction w ith com pu ter.

2 PRESENT WORK Som e system s e.g. Shru ti, Shru ti-II, SAPI etc. are Bengali ASR system . R. Su ltana and R. Palit [17] categorize Bengali STT conversion techniqu es d epend ing on the stu d y topics, alg orithm u sed in d etection and signal analysis m ethod u sed in the recognition system . 1) 2) 3) 4) 5) 6) 7)

Phonem ics and Phonem es Linear Pred iction Cod ing (LPC) Mel Frequency Cepstral Coefficient (MFCC) H id d en Markov Mod el (H MM) Speech Segm entation Acoustic Properties and N eural N etw ork and Fuzzy Logic

2.1 Shruti, Shruti-II S. Mand al, B. Das, P. Mitra [4],[15],[16] presented the versions of Shru ti and Shru ti-II, Bengali ASR System s. Shru ti is a read speech corpu s d esigned to p rovid e speech d ata for the acqu isition of acou stic-phonetic know led ge and for the d evelopm ent and evalu ation of au tom atic speech recognition system s [15]. Shru ti-II, SPH IN X3-based Bengali ASR system converts stan d ard Bengali continu ou s speech to Bengali Unicod e and application for visu ally im paired com m u nity to sen d e-m ail by u sing m ou se and speech as inpu t m ethod . Sphinx-3 is an open sou rce speech recognition engine and u ses 3-state H MM for

IJRCEE@2017 http:/ / w ww.ijrcee.org

International Journal of Research in Computer Engineering and Electronics.

Page # 2

ISSN 2319-376X

VOL :6 ISSUE : 1 (March 2017)

acou stic m od el w hich is triphone based [4], [12]. For state probability d istribu tion it u ses continu ou s d ensity of Gau ssian Mixtu re d istribu tions. All phonem es are m od eled as a sequ ence of H MM state and likelihood of a certain fram e obse rvation is prod u ced by u sing trad itional Gau ssian Mixtu re Mod el. For d eveloping front-end for Shru ti-II w hich integrates E-m ail ap plication and Speech Corpu s Transcripts, Pronu nciation Dictionary, Langu age m od el w hich are based on com m only speaking Bengali Langu age [4]. Shru ti-II contains a total of 7383 u niqu e sentences sp oken by 34 speakers and the speaker age in the corpu s varies in betw een 20 to 40 years only. The accu racy of the Shru ti-II system is abou t 80% for com m only u sed Bengali sentences [4].

2.2 SAPI S. Su ltana, M. A. H . Akhand , P. K. Das, M. M. H . Rah -m an [5] investigated Speech-to-Text (STT) conversion u sing SAPI (Speech Ap plication Program Interface) for Bangla langu age. In this system , first an XML gram m ar file is generated and SAPI [13] context-free gram m ar com piler com piles XML gram m ar into a binary gram m ar form at. The com piled binary gram m ar is load ed into the SAPI ru n tim e environm ent from a file, m em ory, or object (.DLL) resou rce. Then, if a Bangla w ord is spoken, SAPI search it in the binary gram m ar file and retu rns correspond ing pronu nciation of Bangla w ord w ith English characters w hen m atched occu r. The m ethod ology then goes ahead w ith SAPI w ord w ith English character and retu rn a w ord w ith Bangla character fetching a d ictionary. The d ictionary is m anaged w ith a sim ple d atabase so that new w ord ad d ition is sim ple. SAPI u ses inter w ord gap in con -tinu ou s speech for w ord d iscrim ination. If a phrase of spoken w ord s is m isrecognized , it w aits for next phrase to listen. The recognition process continu es u ntil voice phrases enters into the sy stem . Bu t the p roblem of SAPI is that it is a slow process becau se of its sequ ential operations. Also Bangla speech is recognized w ord by w ord basis and a person shou ld speak w ith a proper break in each w ord so that system w rites the w ord if m astch occu rs .So, SAPI converts Bengali speech to text by m atching pronu nciation from continu ou s Bengali speech in precom piled gram m ar file of SAPI and it retu rns Bengali w ord s in English character if m atches occu r.

sonant voice into its correspond ing text in Bengali u sing AN N . Bengali is one of the im portant langu ages w ith a rich heritage. There are tw o types of letters or alp habets [6] in Bengali: vo w el (e.g. অ,আ,ই,ঈ,উ,ঊ,ঋ,এ,ঐ,ও,ঔ) called sorborrn and consonant (e.g. ক, খ ,গ ,ঘ, ঙ, চ, ছ, ম, ঝ, ঞ, ট, ঠ, ড, ঢ, ণ, ট, থ, দ, ধ, ন, ঩, প, ফ, ব, ভ, ম, য, ল,঱, ল, ঳, ঴, ৄ, ৅, য়, ৎ, ং , ং , ং ) called byanjonborrn.We have w orked on consonant voice becau se every consonant voice com bined w ith consonant sou nd and vow el sou nd in Bengali su ch as ক (short consonant sou nd )+ অ (long vow el sou nd ) = ক, ,খ+অ=খ , দ+অ=দ etc.[7]. So vow el sou nd is com m on for all consonant characters and w e d iscard the part of vow el sou nd from the original conconant voice signal. It helps to m atch the inpu t voice signal easily. In ou r propose system , w e u se Artificial N eu ral N etw ork (AN N ) for recognizing the consonant voice. The ad vantage of AN N is its learning process. It can u pd ate the learned inform atin system atically w hile training w ith new inpu t pattern s. Also feed forw ard backprop agation algorithm [9] is u sed to m ake learned the AN N . First, w e take consonant voice from d ifferent people for variation of inpu t voices as sam ple inpu t or training d ata and prosess them . The processing step s involves w ith the convertion of voice signal into vector m atrix, d iscard ing the com m on part in the m atrix and resizing, training AN N u sing backprop agation algorithm and then it w ill be read y for recognition. The processing step s w ill be show n in next section. The pu rpose of system d esign d ocu m ent is to show the logical view of u se case d ia -gram , architectu re d esign, sequ ence d iagram , u ser interface d esign of the system for perform ing the operations su ch as preprocessing au d io, recognize and generalize ou tpu t.

3.1 INFORMATION FLOW DIAGRAM (IFD) Inform ation flow d iagram (IFD) of the prop osed Bengali consonant voice to text conversion system is d ivid ed into tw o sections. First one is Training section and second one is the

3 PROPOSED SYSTEM Technology is changing every d ay. The old one is being replaced by the new one and m od ern technologies are getting m ore popu lar becau se of voice com m and . Du e to the langu age barrier, com m on people face big obstacle to enjoy the optim u m benefits of m od ern com m u nication and inform ation technology (ICT) w here hu ge enriched English know led ge d atabases are there arou nd the globe. Langu age processing in m other tongu e is the only technological w ay that can be u sed to rem ove this barrier. From it, w e are m otivated and aim ed to d evelop su ch a system w hich can convert the Bengali voice into text form . Som e w orks have been d one for som e problem s. H ere, w e d esigned a system w hich can convert the conIJRCEE@2017 http:/ / w ww.ijrcee.org

Fig. 1 Information flow diagram of the proposed system

International Journal of Research in Computer Engineering and Electronics.

Page # 3

ISSN 2319-376X

VOL :6 ISSUE : 1 (March 2017)

sim u lation section w hich is illu strated in Figu re 1.

3.2 USE CASE DIAGRAM

Initialize

Record audio User

Pre-process audio

Cancel

4

EXPERIMENTAL RESULT AND DISCUSSION

Segmentation

We w ent throu gh som e specific step s to com plete this projectlike bu ild ing other system . We d ivid ed the w orking proced u re into tw o sections: 1) Train AN N and 2) Voice Character recognition

Recognize System

Generate output

Fig 2. Use case diagram of the proposed system

The u se case d iagram of the proposed Bengali voice to text conversion system show s the u se case view as in Figu re 2.

3.3 ARCHITECTURE OF THE PROPOSED SYSTEM The architectu re of the proposed Bengali consonant voice to text conversion system consists of som e m od u les illu strated in Figu re 3 and Figu re 4. These are:  Record Au d ios  Inpu t Au d io  Pre-processing  Segm entation and Resize  Training  ecog nition and  u tpu t gen eration Fig. 3 Architecture diagram of the proposed system

4.1 Train ANN The training process of AN N consists of 5 d ifferent steps, these are:  Taking sam ple inpu t voice or training d ata  Convert inpu t au d io into its correspond ing vector m atrix  Preprocessing vector m atrix  Convert vector m atrix into sam ple inpu t m atrix and then

R

Fig. 5 Sample consonant voice input

O  Train AN N

4.1.1 Taking Training Sample We take 20 sam ple inpu t voices from d ifferent peoples for each Bengali consonant character and w e also take voices for sam e character from sam e people for variation. A sam ple voice consonant inpu t signal is show n in Figu re 5.

4.1.2 Convert Input Audio into Its Corresponding Vector Matrix We convert the inpu t voice signal into d ou ble valu ed vector m atrix. Every valu e of vector m atrix is range betw een -1.0 to 1.0. For the above inpu t consonant voice signal the vector m atrix is M=[…… -0.000308 -0.003282 -0.33445435 -0.2342324 …0.20232 0.578387 …. 0.60323…. ].

Fig. 4 Activity diagram of the proposed system

4.1.3 Preprocessing Vector Matrix We processed the d ou ble valu ed vector m atrix into integer IJRCEE@2017 http:/ / w ww.ijrcee.org

International Journal of Research in Computer Engineering and Electronics.

Page # 4

ISSN 2319-376X

VOL :6 ISSUE : 1 (March 2017)

4.2.1 Input Audio We record ed speech from u ser and then take it as a recognition inpu t. We take this inpu t aud io file into variable and variable stores au d io files as vector m atrix valu e w hich is d ou ble valu e in d efau lt ranges from -1.0 to 1.0.

Fig. 6 Sample consonant voice input after discarding

valu ed m atrix. We u sed m apping for conversion and experim enting w ith d ifferent inpu t signal and w e get som e fixed range for all consonant character w hich is the com m on vow el sou nd . So for recognizing the consonant charecter w e d on’t need to m atch the vow el sou nd . That’s w hy w e d iscard the vow el part w hich is show n in Figu re 6.

4.1.4 Convert Vector Matrix into Sample Input Matrix: After that w e have a m atrix w hich contain u n necessary d ata so it need s to resize. While record ing a voice there is an initial gap before and after the original au d io signal w hich actu ally is the preparation for d elivering voice by u ser. We d iscard the signal and also need to d iscad the inpu t m atrix. So w e get the m atrix M after d iscard ing as M = [.......0.0318 0.03282 0.33445435 0.2342324……0.20232 0.578387……0.60323……]. 4.1.5 Train ANN We u sed su pervised learning to train AN N u sing the au tom ated feed forw ard back propagation algorithm tool of MATLAB 7.14.0.334 [8]. We set 390 sam ples of inpu t voice characters and each character of 10 d ifferent sam ples. We train this AN N only for first 10 Bengali characters and set the nu m bers 1 to 10 for them . A key featu re of neu ral netw orks is an iterative learning process in w hich d ata cases (row s) are presented to the netw ork one at a tim e, and the w eights associated w ith the inpu t valu es are ad ju sted each tim e. The training continu ed for 6 m ore iteration before the training stop ped . After all cases are presented , the process often starts over again. Du ring this learning phase, the netw ork learns by ad ju sting the w eights so as to be able to pred ict the correct class label of inpu t sam ples [19].

4.2 Voice Character Recognition We d ivid ed this section into som e step s:  Inpu t au d io (for recognition)  Process au d io  Recognition

4.2.2 Process Audio When w e record voice then there exist initial gap w hich is n ot a part of m ain voice. And w hen the consonant characters are pronou nced it contains a vow el part in its last. And so to re cognize a consonant character it m ay cau se no effect if w e d iscard the vow el part from main voice. Au d ios are continu ou s signals. It can be of variou s lengths. 4.2.3 Recognition AN N gave approxim ate ou tpu t resu lt accord ing to the learned target inpu t to it. And for the ou tpu t w e m atched it w ith ou r all target and choose the short d istance valu e for it and it is ou r ou tpu t for recognition. Pseu d o cod e for recognition: minValue = infinity approximate-value = AN N (inp u t) for Each x = target valu e for character dist = absolu te(x - approximate-value) if dist < minValue ans = x and minValue = dist return ans.

4.3 Experimental Result Cross-valid ation [20] is a com m on and effective testing proced u re. To test this proposed system ten-fold cross-valid ation is ap plied . In a ten-fold cross-valid ation, it is requ ired to split the w hole training set into su bsequ ently, one su bset is tested u sing the classifier trained on the rem aining nine su bsets. As the testing su bset is u nknow n to the classifier, the su ccess rate of the pred iction accu racy obtained from the u nknow n su bset. Therefore, the testing proced u re is able to prevent over -fitting and the resu lt generalizes better to the actu al operating env ironm ent [18]. 4.4 Performance and Accuracy We first choose som e au d io sam ple. After that each featu re is tested ind epend ently for d im ensionality red u ction. Then overall system perform ance is observed . To select best m atch character, w e train the AN N only w ith featu res for d ifferent nu m ber of sam ple au d io of d ifferent character and it gives satisfactory perform ance. If the training w ere perfect, the netw ork ou tpu ts and the targets w ou ld be exactly equ al, bu t the relationship is rarely perfect in practice. Ou r proposed system requ ires less m em ory space com pared to the others becau se of the d im entionality red u ction

IJRCEE@2017 http:/ / w ww.ijrcee.org

International Journal of Research in Computer Engineering and Electronics.

Page # 5

ISSN 2319-376X

VOL :6 ISSUE : 1 (March 2017)

4.10 Test Results The percentage of recognition accu racy of voice consonent characters is 66.67% w here 10 training sam ples and 5 testing sam ples are u sed for learning and recognizing for each Bengali consonant character. The testing ou tpu ts are show n in Table 1.

TABLE 1 RECOGNITION RESULT AFTER TRAINED ANN Characters

Training Samples

Testing Sample

Correctly Recognized



10

5

5

%of Recognition Accuracy 100%



10

5

5

100%



10

5

5

100%



10

5

5

100%



10

5

1

20%



10

5

3

60%



10

5

4

80%

জ ঝ

10 10

5 5

5 4

100% 80%



10

5

1

20%



10

5

4

80%



10

5

4

80%

ড ঢ

10 10

5 5

5 5

100% 100%

5 CONCLUSION Speech recognition (and then conversion) cannot be error free. Environm ental cond itions, the style of speech effects on the speech transcript accu racy w hich is highly d epend ent on the speaker. Speech recognition is a hard er process than w hat people com m only think, even for a hu m an being. H um ans are u sed to u nd erstand ing speech, not to transcribing it, and only speech that is w ell form u lated can be transcribed w ithou t am bigu ity. Hu m an brain is far too pow erfu l and this shou ld not be com pared w ith ou r system . H u m an brain is also learned for recognizing voice. As ou r system is trained u sing su pervised learning, it w ill be m ore pow erfu l if w e train it m ore and m ore. In this project, w e presented a robu st and com pu tatio nally efficient m ethod for conversion of Bengali consonant voice character to text.

5.1 Limitations and Future Plan Most of the tim e the perform ance of the neu ral netw ork increases w ith the increase of training d ata. So enou gh training d ata is need ed to train neu ral netw ork. If noisy sam ple in pu ts are taken, then it can’t give the correct ou tpu t. We have w orked here on Bengali character bu t in fu tu re w e’ll also w ork on w ord s as w ell as sentence for recognition. We w ill m ake the system to w ork for noisy inpu t. More training exam ples w ill be u sed to train the system .



10

5

5

100%



10

5

5

100%



10

5

5

100%

REFERENCES



10

5

5

100%

[1]



10

5

4

80%



10

5

4

80%



10

5

3

60%



10

5

3

60%



10

5

5

100%



10

5

5

100%



10

5

3

60%



10

5

4

80%



10

5

3

60%



10

5

4

80%



5

5

100%



10 10

5

0

0%



10

5

0

0%

[2]

[3]

[4]

[5]



10

5

3

60%



10

5

3

60%



10

5

0

0%



10

5

1

20%

[6]



10

5

0

0%



10

5

1

20%



10

5

1

20%



10

5

2 Total

40% 66.67%

[7] [8]

N. Deshmukh, A. Ganapathiraju and J. Picone, “Hierarchical Search for Large Vocabulary Conversational Speech Recognition,” IEEE Signal Processing Magazine, vol. 1, no. 5, pp. 84-107, September 1999. M. Jackson, “Automatic Speech Recognition: H uman Com puter Interface for Kinyarw and a Language”, Master Thesis, Faculty of Com puting and Information Technology, Makerere University, 2005. V. K. Singh, “Most Spoken Languages in the world ”, http:/ / w ww.listsw orld .com / top -10-languages-m ost-spokenw orld w id e. 2012. S. Mandal, B. Das, P. Mitra, “Shruti-II: A vernacular speech recognition system in Bengali and an application for visually impaired community”, Student’s Technology Symposium (TechSym), 2010 IEEE Conference, pp. 229 – 233, April 3-4, 2010. S. Sultana, M. A. H. Akhand, P. K. Das, M. M. Hafizur Rahman, "Bangla speech-to-text conversion using sapi", Computer and Communication Engineering (ICCCE) 2012 International Conference on, pp. 385-390, July 3-5, 2012. M. M. Hossain, A. Habib, M. S. Rahman, “Transliteration Based Bengali Text Compression Using Huffman Principle”, International Conference on Informatics, Electronics & Vision (ICIEV), pp-1, 2014. Unnam ed , “ধ্বনন ও ফণণ প্রকযণ ও উচ্চাযণনফনধ” Available at http:/ / w ww.ed pd bd .org/ uap/ bangla/ ধ্বনন-ও-ফণণ-প্রকযণ-ও-উচ্চাযণনফনধ. Gonzalez, R.C., R.E., “Digital Image Processing. Using MATLAB”, Chap: 2, pp. 1-30, 1992.

IJRCEE@2017 http:/ / w ww.ijrcee.org

International Journal of Research in Computer Engineering and Electronics.

Page # 6

ISSN 2319-376X

VOL :6 ISSUE : 1 (March 2017)

[9] [10]

[11]

[12]

[13]

[14]

[15]

[16]

[17]

[18] [19]

Unnamed, “Training an Artificial Nseural Network – Intro”, Available at http://www.solver.com/training-artificial-neural-network-intro.2012. Unnamed, “Analyze neural network performance after training”, Available https://www.mathworks.com/help/nnet/ug/analyze-neural-networkperformance-after-training.html.2016. S. Patel, A. Bramhecha, S. Mahale, A. Maind, M. Sanghavi,“Speech Recognition System for Windows Commands”, International Journal of Computer Applications International Conference on Recent Trends in engineering & Technology (ICRTET'), pp-31,2013. S. Darjaa, M. Cerňak, M. Trnka, M. Rusko, R. Sabo, "Effective triphone mapping for acoustic modeling in speech recognition," In Proceedings of Interspeech, pp. 1717-1720, 2001. Unnamed, “Speech Application Program Interface (SAPI)”, Available at http://whatis.techtarget.com/definition/Speech-Application-ProgramInterface-SAPI. 2005. P. Barua, K. Ahmad, A. A. S. Khan, M. Sanaullah, “Neural network based recognition of speech using MFCC features”, International Conference on Informatics, Electronics & Vision (ICIEV), 2014. B. Das, S. Mandal, P. Mitra, "Bengali speech corpus for continuous automatic speech recognition system," Proc. Conf. Speech Database and Assessments (Oriental COCOSDA), pp.51-55, 2011. S. Mandal, B. Das, P. Mitra, A. Basu, "Developing Bengali Speech Corpus for Phone Recognizer Using Optimum Text Selection Technique," Proc. Conf. Asian Language Processing (IALP), 2011. R. Sultana and R. Palit, “A survey on Bengali speech -to-text recognition techniques”, IEEE 9th International Forum on Strategic Technology (IFOST) 2014. Unnamed, “3.1.Cross-validation: evaluating estimator performance”, Available at http://www.openml.org/a/estimation-procedures/1. 2010-2016. M. Cilimkovic, “Neural Networks and Back Propagation Algorithm”, Available at http://www.dataminingmasters.com/uploads/studentProjects/NeuralNetwork s.pdf. 2010. [20] Unnamed, “Cross Validated”, Available at http://stats.stackexchange.com/questions/27730/choiceof-k-in-k-fold-cross-validation. 2015.

Author’s Biography Shikhor Kumer Roy (shikhorroy.cse12@gm ail.com ) received his B.Sc. Degree from H ajee Moham m ad Danesh Science and Technology University, Dinajpur, Bangslad esh in 2015. H e likes to solve problem s in Java and C++ m ostly. H e participated in m any program m ing contests and achieved good w ill at the tim e of grad uation. H is another tw o w orks of Data m ing is published on 2017 IEEE International Conference on Im aging, Vision and Pattern Recognition and an Am erican open access journal (AJER). H e is m ainly a softw are d eveloper (program m er). Abinash Kanti Ghosh (abinashkg@gm ail.com ) received his B.Sc. d egree in Com puter Science and Engin eering from H ajee Mo-ham m ad Danesh Science and Technology University, Dinajpur, Banglad esh in 2015. H is research interests in Voice Recognition, Artificial Intelligence and Machine Learning. Aysa Siddika Asa (asha.cse12@gm ail.com ) received her B.Sc. d egree in Com puter Scienge and Engineering from H ajee Mo-

ham m ad Danesh Science and Technology University, Dinajpur, Bangslad esh in 2015. She is very active, hard w orking, dedicated and analytical person. H er interested research areas are Data Mining, Im age Processing, Cloud com puting Machiine learning, Artificial Intelligence etc. She also has another tw o publications on the behalf of final year thesis on Data Mining. Md. Palash Uddin ([email protected] ) received his B.Sc. d egree in Com puter Science an d Engineering from H ajee Moham m ad Danesh Science and Technology University, Dinajpur, Banglad esh in 2011 (H eld in 2013). H is m ain w orking interest is based on artificial intelligence based application d evelopm ent, m achine learning, algorithm analysis, d ata m ining and im age analysis. Currently, he is w orking as an assistant professor in d ept. of Com puter Science and Engineering in H ajee Moham m ad Danesh Science and Technology University, Dinajpur, Banglad esh. Previously, he w as a lecturer in the sam e university and in d epartm ent of Com puter Science and Engineering at Central Wom en ’s University, Dhaka, Banglad esh. He has research several publications in various field s of Com pu ter Science and Engineering. Md. Rashedul Islam (rashed [email protected] ) received his B.Sc. d egree in Com puter Science and Engineering from H ajee Moham m ad Danesh Science and Technology University, Dinajpur, Banglad esh in 2012 (H eld in 2014). H is m ain w orking interest is based on d ata m ining, im age processing and security techn iques. Currently, he is w orking as a lecturer in d ept. of Com puter Science and Engineering in H ajee Moham m ad Danesh Science and Technology University, Dinajpur, Banglad esh. Previously, he w as a lecturer in The University of Developm ent Alternative (UODA) and Uttara University, Banglad esh. H e has research publications in d ifferent field s of Com puter Science and Engineering. Masud Ibn Afjal (m asud @hstu.ac.bd ) received his B.Sc. d egree in Com puter Science and Engineering from H ajee Moham m ad Danesh Science an d Technology University, Dinajpur, Banglad esh in 2008. H is m ain w orking interest is based on m achine learning, algorithm analysis, d ata m ining, im age analysis and d atabase. Currently, he is w orking as an assistant professor in d ept. of Com puter Science and Engineering in H ajee Moham m ad Danesh Science and Technology University, Dinajpur, Banglad esh. Previously, he w as a lecturer in the sam e university and also served in various softw are d evelopm ent com panies of Banglad esh. H e has research publications in d ifferent field s of Com puter Science and Engineering.

IJRCEE@2017 http:/ / w ww.ijrcee.org

International Journal of Research in Computer Engineering and Electronics. VOL :6 ISSUE : 1 (March 2017)

IJRCEE@2017 http:/ / w ww.ijrcee.org

Page # 7

ISSN 2319-376X