A lightweight speech recognition method with target-swap knowledge distillation for Mandarin air traffic control...

Area of Science:

• Aviation safety
• Speech and language processing
• Machine learning

Background:

• Miscommunications between air traffic controllers (ATCOs) and pilots in air traffic control (ATC) pose significant safety risks.
• Automatic speech recognition (ASR) offers a promising solution to mitigate these misunderstandings.
• Existing ASR systems for ATC often prioritize recognition accuracy over transcription speed, creating a critical gap.

Purpose of the Study:

• To enhance the performance and reduce the transcription latency of ASR systems for Mandarin ATC communications.
• To introduce a novel knowledge distillation strategy for developing lightweight ASR models.
• To improve the balance between recognition accuracy and transcription speed in ATC ASR.

Main Methods:

• Introduction of Target-Swap Knowledge Distillation (TSKD), a lightweight strategy for ASR.
• TSKD involves swapping logit outputs for the target class between teacher and student models.
• Application of TSKD to Mandarin ATC communications within homogeneous and heterogeneous architectures.

Main Results:

• The proposed TSKD strategy effectively enhances the generalization performance of lightweight ASR models.
• Experimental results demonstrate the effectiveness of TSKD in improving the balance between accuracy and latency.
• The developed lightweight ASR model achieves superior performance in both recognition accuracy and transcription speed.

Conclusions:

• TSKD is a simple yet effective method for developing high-performance, low-latency ASR systems for ATC.
• The research addresses the critical need for faster ASR systems in time-sensitive aviation environments.
• The findings contribute to improving aviation safety by reducing communication errors through advanced ASR technology.