ATCSpeechAgent: A Few-Shot Speaker-Adaptive Text-to-Speech Framework for Pseudo-Pilot Generation in Mandarin Air Traffic Control

Realistic and acoustically varied pilot speech is essential for effective air traffic controller (ATCO) training. However, acquiring sufficient real-world recordings is costly and logistically challenging, particularly in low-resource settings. This paper presents ATCSpeechAgent, a speaker-adaptive text-to-speech framework for generating high-fidelity pseudo-pilot utterances in Mandarin air traffic control (ATC) simulations. Leveraging a few-shot learning paradigm, the framework achieves accurate voice cloning from only a handful of samples. A hierarchical acoustic condition modeling architecture is designed to encode and integrate speaker-, utterance-, and phoneme-level cues, thereby enhancing generalization to diverse acoustic scenarios and unseen speakers. Building on this foundation, a meta-learning-based speaker adaptation strategy establishes speaker-invariant model initialization, enabling rapid and robust adaptation to novel pilot voiceprints. Extensive experiments on Mandarin ATC and open-domain corpora show that the proposed approach consistently outperforms competitive baselines in speaker similarity and naturalness under few-shot conditions, highlighting its potential for scalable, resource-efficient, and operationally realistic ATCO simulation training.