Text-To-Speech Application for Simulation of Aviation Radio Telephony Communication Partners

Kurzfassung

Air Traffic Control (ATC) is a critical component of aviation safety, ensuring efficient flow of air traffic without accidents. In the wake of the COVID-19 pandemic, the aviation industry faced a sudden shortage of pilots and Air Traffic Controllers (ATCOs), highlighting a significant problem: managing resources for training new ATCOs and pilots. In response to this challenge, this thesis explores the concept of using a Text-toSpeech (TTS) application to simulate aviation radio telephony communication. The application utilizes open-source pre-trained TTS models fine-tuned using publicly available ATC communication-specific datasets. It synthesizes textual ATC utterances to simulate ATCO instructions and pilot responses, creating a realistic two-way communication scenario. It includes twenty ATCO and eight pilot voice models developed using two distinct fine-tuning approaches. The first approach being an end-to-end TTS method that leverages advanced deep learning techniques, while the other is a voice cloning method that supports multi-lingual speech generation. These different approaches led to significant quality variations in the resulting voice models. Real-Time Factor (RTF) scores of less than 1 were achieved, demonstrating faster than real-time processing, when running the application with required computational resources. The application was evaluated by a group of 14 international individuals, comprising 4 pilots, 8 ATCOs, and 2 testers from other aviation backgrounds. The performance of the models varied across different aspects of audio quality such as overall score, naturalness, clarity, and pronunciation. The highest rated ATCO and pilot models received average overall quality scores of 4.5 and 4.2 respectively out of highest score of 5. The average overall quality rating for all ATCO models was approximately 4.0, while for pilot models it was 3.6 out of 5. For the ATCO and pilot models, 86.7% and 84.4% of the ratings respectively classified the speed of the generated speech as ‘Optimal’. Through individual comments on the audio output, testers appreciated the realism of some voice models, indicating their potential for effective training simulations. However, for some models, they noted issues with pronunciation, pace, and volume consistency, emphasizing the need for clear numerical pronunciation and correct phraseology. Overall user feedback shows that the application has potential to contribute to ATCO/pilot training by providing a cost-effective and efficient tool for simulating ATC communication, thereby enhancing the training experience for aviation radio telephony communication partners. Future works could focus on integrating advanced text generation models like GPT- 4 and open-source models like LLaMA, fine-tuned with ATC-specific data for dynamic, contextual responses, which would be a significant improvement over the current approach which uses hard-coded regular expressions for generating pilot response. Additionally, incorporating real-world noise into pilot communications could enhance realism.