Prosodically Guided Phonetic Engine for Punjabi Language

Abstract

This thesis deals with prosodically guided phonetic engine for Automatic Speech Recognition (ASR) for Punjabi language using trainable systems. The goal of this work is the development of a phonetic Engine using Phonetic Transcription and so to build acoustic models for Punjabi language. This is done by employing Hidden Markov Models (HMMs) that provide statistical representation of each of the distinct sounds that make up a word and to train the parameters of the models developed. The speech recognition modelling can be done in two ways: Acoustic modelling and Language Modelling. Out of these two, Acoustic Modelling has been used. In this work, Acoustic modelling has been worked out at a phonetic level, allowing general speech recognition applications. For this purpose, the tool HTK is employed, stated as Hidden Markov Model toolkit. HTK uses different commands to produce HMM models for each phone and computing various parameters for mixture. Different phone models have been developed and tested. This thesis is divided into six chapters. A brief review of these chapters is given below.

Chapter 1 includes the definitions of the terms and the brief idea of the concepts, description of tools used in this work.

Chapter 2 includes the literature survey which depicts the ideology, and previously proposed models and methodologies to train a phonetic engine.

Chapter 3 depicts the problem statement and its cause that motivates to work in this domain.

Chapter 4 describes the process of data collection of read, lecture, and conversational modes of Punjabi Speech, experimental set up and methodology which depicts the process of development of prosodically guided phonetic engine. For this purpose, Ubuntu Linux (14.04 64 bit) has been considered as environment. Wave Surfer tool has been used for manual transcription and Hidden Markov Model toolkit has been used to train and test the phonetic engine for various modes of speech.

Chapter 5 presents the results of phonetic engine developed in this work for two categories, 30 phones and 34 phones. Gender-wise and transcription-wise performance of the developed engine have also been presented in this chapter.

Chapter 6 includes the conclusion and future scope of this work.