A Comprehensive Review on Phoneme Classification in ML Models

Abstract

This paper gives a relative performance examination of both shallow and profound machine learning classifiers for speech recognition errands utilizing outline level phoneme classification. Phoneme recognition is as yet a principal and similarly significant introductory advance toward automatic speech recognition (ASR) frameworks. Frequently regular classifiers perform outstandingly well on domain-explicit ASR frameworks having a constrained arrangement of jargon and preparing information as opposed to profound learning draws near. It is consequently basic to assess the performance of a framework utilizing profound artificial systems regarding effectively perceiving nuclear speech units, i.e., phonemes right now customary cutting-edge machine learning classifiers. Two profound learning models - DNN and LSTM with numerous arrangement structures by changing the quantity of layers and the quantity of neurons in each layer on the OLLO speech corpora alongside with six shallow machines get the hang of ing classifiers for Filterbank acoustic features are completely considered. Moreover, features with three and ten edges transient setting are registered and contrasted and no-setting features for various models. The classifier's performance is assessed as far as accuracy, review, and F1 score for 14 consonants and 10 vowels classes for 10 speakers with 4 distinct tongues. High classification precision of 93% and 95% F1 score is gotten with DNN and LSTM organizes separately on setting subordinate features for 3-shrouded layers containing 1024 hubs each. SVM shockingly acquired even a higher classification score of 96.13% and a misclassification blunder of under 5% for consonants and 4% for vowels.