Title: Boosting speech recognition performance: a robust and accurate ensemble method based on HMMs

Authors: Samira Hazmoune; Fateh Bougamouza; Smaine Mazouzi; Mohamed Benmohammed

Addresses: Department of Computer Science, Faculty of Sciences, University of 20 Août 1955-Skikda, Skikda, Algeria ' Department of Computer Science, Faculty of Sciences, University of 20 Août 1955-Skikda, Skikda, Algeria ' Department of Computer Science, Faculty of Sciences, University of 20 Août 1955-Skikda, Skikda, Algeria ' Department of Software Technologies and Information Systems, Faculty of New Technologies of Information and Communication, University Constantine 2, Constantine, Algeria

Abstract: In this paper, we propose an ensemble method based on hidden Markov models (HMMs) for speech recognition. Our objective is to reduce the impact of the initial setting of training parameters on the final model while improving accuracy and robustness, particularly in speaker independent systems. The main idea is to exploit the sensitivity of HMMs to the initial setting of training parameters, thus creating diversity among the ensemble members. Additionally, we perform an experimental study to investigate the potential relationship between initial training parameters and ten diversity measures from literature. The proposed method is assessed on a standard dataset from the UCI machine-learning repository. Results demonstrate its effectiveness in terms of accuracy and robustness to intra-class variability, surpassing basic classifiers (HMM, KNN, NN, SVM) and some previous works in the literature including those using deep learning algorithms such as convolutional neural networks (CNNs) and long short-term memory (LSTM).

Keywords: speech recognition; inter-speaker variability; robustness; accuracy; HMM; multiple modelling; ensemble methods; diversity.

DOI: 10.1504/IJISTA.2024.136523

International Journal of Intelligent Systems Technologies and Applications, 2024 Vol.22 No.1, pp.41 - 76

Received: 13 Jul 2023
Accepted: 19 Sep 2023
Published online: 05 Feb 2024 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article