Article: Optimisation of English speech recognition using contrastive learning and feature reconstruction Journal: International Journal of Information and Communication Technology (IJICT) 2026 Vol.27 No.98 pp.1 - 25 Abstract: Speech recognition still faces challenges such as insufficient feature discriminability, severe noise interference, and low training efficiency in complex noise and multi accent scenarios. The existing English speech recognition methods mainly suffer from fuzzy feature discrimination boundaries, coarse-grained noise robustness modelling, limited cross accent generalisation ability, and high dependence on annotated data and slow convergence speed during model training. This study proposes a model integrating enhanced contrastive training, feature reconstruction, and parameter optimisation. It introduces a momentum contrast hybrid algorithm to reduce dependence on negative sample quality in traditional contrastive learning, combined with feature reconstruction for better information integrity. A deep neural network-ideal ratio mask module optimises noise suppression. These components are fused for global parameter optimisation. On benchmark datasets, it achieves a feature reconstruction error as low as 0.113, a heterogeneous-to-homogeneous distance ratio up to 3.56, and a maximum SNR improvement of 16.2 dB. Furthermore, it attains a recognition accuracy of 98.23%, a word error rate of 1.03%, and a short-term objective intelligibility score of 0.962. This provides robust technical support for real-time, high-accuracy applications like intelligent assistants and real-time translation. Inderscience Publishers - linking academia, business and industry through research

Title: Optimisation of English speech recognition using contrastive learning and feature reconstruction

Authors: Jing Ren

Addresses: School of International Education, Changchun Finance College, Changchun, 130000, China

Abstract: Speech recognition still faces challenges such as insufficient feature discriminability, severe noise interference, and low training efficiency in complex noise and multi accent scenarios. The existing English speech recognition methods mainly suffer from fuzzy feature discrimination boundaries, coarse-grained noise robustness modelling, limited cross accent generalisation ability, and high dependence on annotated data and slow convergence speed during model training. This study proposes a model integrating enhanced contrastive training, feature reconstruction, and parameter optimisation. It introduces a momentum contrast hybrid algorithm to reduce dependence on negative sample quality in traditional contrastive learning, combined with feature reconstruction for better information integrity. A deep neural network-ideal ratio mask module optimises noise suppression. These components are fused for global parameter optimisation. On benchmark datasets, it achieves a feature reconstruction error as low as 0.113, a heterogeneous-to-homogeneous distance ratio up to 3.56, and a maximum SNR improvement of 16.2 dB. Furthermore, it attains a recognition accuracy of 98.23%, a word error rate of 1.03%, and a short-term objective intelligibility score of 0.962. This provides robust technical support for real-time, high-accuracy applications like intelligent assistants and real-time translation.

Keywords: feature comparison training; FCT; feature reconstruction; FR; supervised contrastive learning; SCL; algorithm optimisation; English speech recognition.

DOI: 10.1504/IJICT.2026.156258

International Journal of Information and Communication Technology, 2026 Vol.27 No.98, pp.1 - 25

Received: 20 Jan 2026
Accepted: 12 Jun 2026
Published online: 08 Sep 2026 *

Title: Optimisation of English speech recognition using contrastive learning and feature reconstruction

Keep up-to-date