Optimizing feature extraction techniques constituting phone based modelling on connected words for Punjabi automatic speech recognition

Abstract

Punjabi phoneme sounds are tonal in nature which dissent in most regions of Punjab. Recent research works reveal less significant work done towards developing a speech recognition system in Punjabi. The work done will feature out variability in the correctness and accuracy of various feature extraction techniques. Following paper objects the application of Automatic Speech Recognition on connected words instituting HTK toolkit modelled on Hidden Markov Model (HMM) to build the system. Back-end of the system was braced for 150 distinct Punjabi words from 16 distinct speakers from noise-free corpus and 12 speakers were indulged for the collection of noisy corpus including both male and female. In the phrase of speech recognition, the proposed Feature Extractor we use Front end techniques as “power normalized cepstral coefficients (PNCC)”, “Mel frequency cepstral coefficients (MFCC)” and “Perceptual Linear Prediction (PLP)” following a statistical comparison based on the accuracy and correctness of results attained. To attain a higher rate of accuracy level 34 phones for Punjabi language are used to break each word into small sound frames. Hence, the comparison based on the nature of training and testing environment will aid in framing a vital speech recognition system for Punjabi language.