A Unified Framework for Multilingual Speech Recognition in Air Traffic Control Systems

Area of Science:

• Speech Processing
• Artificial Intelligence
• Aviation Technology

Background:

• Air traffic control (ATC) demands highly accurate speech recognition systems.
• Existing systems struggle with radio noise and multilingual communication.

Purpose of the Study:

• To develop a robust, multilingual speech recognition framework for ATC.
• To improve accuracy and efficiency in processing ATC communications.

Main Methods:

• A novel cascaded framework with acoustic (AM), pronunciation (PM), and language models (LM).
• Utilized convolutional neural networks (CNN) and recurrent neural networks (RNN) for AM.
• Implemented a machine translation encoder-decoder architecture for the PM.
• Developed RNN-based LMs to handle code-switching in ATC speech.

Main Results:

• Achieved a 3.95% label error rate on Chinese characters and English words.
• Outperformed existing popular speech recognition approaches.
• Demonstrated comparable decoding efficiency to end-to-end models.
• Validated generalizability on open corpora.

Conclusions:

• The proposed cascaded model offers robust and efficient multilingual speech recognition for ATC.
• The system's performance and generalizability support real-time applications in ATC.
• This framework can enhance ATC safety and prediction capabilities.