Work

Sentence-level consistency of conformer based pre-training distillation for Chinese speech recognition

Year: 2025

Type: article

Abstract: Introduction We address robustness and efficiency in Chinese automatic speech recognition (ASR), focusing on long-form broadcast speech where sentence-level semantic consistency is often lost. Methods We propose a Conformer-based framework that integrates sentence-level consistency with pre-training knowledge distillation. We also construct CH Broadcast ASR, a domain-specific Chinese corpus for the broadcast and television domain, and evaluate on AISHELL-1, AISHELL-3, and CH Broadcast ASR. Results The proposed model consistently outperforms strong baselines (TDNN, DFSMN-T, TCN-Transformer), achieving CER = 3.3% on AISHELL-1, 3.7% on AISHELL-3, and 3.9% on CH Broadcast ASR, while reducing model size by >10%. Discussion Enforcing sentence-level semantic alignment together with distillation improves robustness for long-form broadcast speech and enhances efficiency for real-time deployment.

Source: Frontiers in Communications and Networks

DOI: 10.3389/frcmn.2025.1662788

Authors: Haifang Li, Chao Tang, Xin Yue, Xu Li

Institution: Xinjiang Normal University

Language: English

Topic: Speech Recognition and Synthesis

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: gold

Estimated APC paid: $2,353

Funder: Xinjiang Normal University