Showing papers by "Santanu Phadikar published in 2015"

Writer Identification from Handwritten Devanagari Script

Abstract: This paper presents analysis of Devanagari characters for writer identification. Being originated from Brahmic script, Devanagari is the most popular script in India. It is used by over 400 million people around the world. Application of writer identification of Devanagari handwritten characters covers a vast area such as The Questioned Document Examination (QDE) is an area of the Forensic Science with the main purpose to answer questions related to questioned document (authenticity, authorship and others). Signature verification in banking, in Graphology (study of handwriting) a theory or practice for inferring a person’s character, disposition, and attitudes from their handwriting. Here we collect 5 copies of handwritten characters to nullify intra-writing variation, from 50 different people mainly students. After preprocessing and character extraction, 64-dimensional feature is computed based on gradient of the images. Some manual processing is required because some noises are too difficult to remove automatically as they are much closer to the characters. We have used LIBLINEAR and LIBSVM classifiers of WEKA environment to get the individuality of characters. We have done the writer identification with all the characters and obtained 99.12 % accuracy for LIBLINEAR with all writers. Features collected from this work can be used in the next level to identify writers from their cursive writing.

Modified energy based method for word endpoints detection of continuous speech signal in real world environment

Abstract: Accurately identifying the word endpoints is an important step of speech recognition process. This paper proposes a robust word endpoints detection algorithm of continuous speech signal collected from real world environment. In this process energy feature is used along with zero crossing rate feature to locate the endpoints of word in speech signal. A set of 100 different sentences have been recorded from 10 speakers which are used as referred dataset. Proposed method is applied on that dataset and accuracy has been measured by computing the average difference between ground truth endpoints (manually estimated) and system generated endpoints. This algorithm attains 85.5% accuracy whereas the entropy based method gives the accuracy of 78.6 % which shows the superiority of the proposed method. Euclidean distance and Manhattan distance for the proposed algorithm is 2.4 and 2.8 respectively which is also quite acceptable.

Automatic Segmentation of Spoken Word Signals into Letters Based on Amplitude Variation for Speech to Text Transcription

Abstract: In this paper a technique for automatic segmentation of spoken word signals is presented for identifying letters for transcription into textual form. Signal patterns for each letter present in different words have been used for the purpose. Voice signals are obtained by taking pronunciations of 1,000 words available in the standard dictionary. After collecting the signals, pre-processing is performed to reduce the noise taking a heuristically determined threshold value. Then the signals are segmented based on Amplitude Variation (AV) in different portions of the signal, each corresponding to an alphabet in that particular word. Signal Peak Value (SPV) is the feature used for recognizing the letters. Accuracy of the method is estimated using Bagging, Bayes Net, J48, Naive Bayes, PART and SVM classifiers available in Weka. The best and the average classification accuracies obtained in this method are 95.15 % (given by J48 classifier) and 86.92 %, respectively, which are quite acceptable.