A comparative study of classification techniques for English to Arabic speech recognition

Other Title(s)

دراسة مقارنة لتقنيات التصنيف للتعرف على ترجمة الكلام من اللغة الإنجليزية إلى اللغة العربية

Dissertant

al-Umari, Ayyub Abd al-Rahman

Thesis advisor

Nasir al-Din, Hibah Hasan O.

Comitee Members

al-Husayni, Muhammad Abbas Fadil
Tani, Ahmad

University

Middle East University

Faculty

Faculty of Information Technology

Department

Department of Computer Information Systems

University Country

Jordan

Degree

Master

Degree Date

2016

English Abstract

Speech processing is considered to be one of the most important application area of digital signal processing.

Speech recognition and translation systems have consisted into two main systems, the first system represents an ASR system that contains two levels which are level one the feature extraction level As well as, level two the classification technique level using Data Time Wrapping (DTW), Hidden Markov Model (HMM), and Dynamic Bayesian Network (DBN).

The second system is the Machine Translation (MT) system that mainly can be achieved by using three approaches which are (A) the statistical-based approach, (B) rule -approach, and (C) hybrid-based approach.

In this study, we made a comparative study between classification techniques from ASR point of view, as well as, the translation approaches from MT point of view.

The recognition rate was used in the ASR level and the error rate was used to evaluate the accuracy of the translated sentences.

Furthermore, we classified the sample text audio files into four categories which were news, conversational, scientific phrases, and control categories.

The empirical findings showed that the DBN achieved the best recognition rate for news category with 79.2% compared with HMM and DTW.

However, the HMM classification technique achieved the highest accuracy in term of recognition rate for conversational with 80.1%, scientific phrases with 86%, and control with 63.8 % recognition rates.

In contrast, using DTW in ASR had a negative behavior on the recognition rate for all speech categories.

The rule-based model which was represented by IBM Watson cloud achieved high translation accuracy results for the majority of speech categories with 13.93% in conversational, 7.38% in scientific phrases, and 17.91% in control categories.

However, by using the statistical-based model – that was represented by Google Translate - in translation the empirical findings showed that for conversational and scientific phrases the error rate was close to rule based with an intangible difference.

In contrast, by using the hybrid-based model influenced the error rate in the three ASR classification techniques and for all speech categories which was assigned as a negative effects.

Main Subjects

Information Technology and Computer Science

No. of Pages

99

Table of Contents

• APA
• MLA
• AMA

Language

English

Data Type

Arab Theses

Record ID

BIM-721130