A Convolutional Hybrid Models of Parametrization in Speech Recognition

Abstract

This research evaluates a novel convolutional parametrization model that integrates feature vectors derived from the Gammachirp Wavelet Transform into traditional parametrization frameworks. It benchmarks the performance of the proposed hybrid parameters (GWTFCC-GWTPLP) against conventional methods, including Mel-Frequency Cepstral Coefficients (MFCC) and Perceptual Linear Prediction Coefficients (PLP), within an isolated-word automatic speech recognition (ASR) system. The aim of this study is to demonstrate the impact of optimized hybrid parameters on recognition accuracy under varying training and testing conditions, particularly in noisy environments. The evaluation focuses on the robustness of the extracted features and their contribution to system effectiveness.