This study focuses on a deep learning-based automatic announcement detection system for passenger English, aiming to improve the accuracy and real-time interactivity of speech recognition in air travel. The system architecture includes a front-end signal processing, a back-end deep learning model and a middleware layer, which purifies the signal through noise suppression, echo cancellation, and speech activity detection, and extracts and transforms the signal features by using the AABLT model, an adaptive attention mechanism model combining bidirectional LSTM and Transformer. We constructed a high-quality dataset covering comprehensive broadcast content and complex noise environments, and achieved highly accurate speech recognition with the AABLT model. Experimental evaluations show that the AABLT model exhibits excellent recognition performance and robustness under different noise levels and accent conditions, significantly reducing word error rate and sentence error rate while optimizing resource consumption and response time. This work addresses a safety-critical gap in aviationreliable, low-latency detection and recognition of crew English announcements under cabin noise and accent variability. We design an AABLT architecture—an adaptive-attention BiLSTM-Transformer—with a purpose-built dataset covering operational phases and heterogeneous accents, and a middleware that enables real-time interaction onboard. Compared with strong baselines, our system achieves materially lower WER/SER and faster responses in streaming conditions, yielding up to about 25%–30% relative improvements under representative SNRs while reducing GPU/CPU utilization. Beyond accuracy, we contribute a deployable pipeline that integrates front-end denoising/VAD, lowlatency decoding, and context-aware delivery, providing actionable guidance for airlines to enhance announcement reliability and passenger safety. This research provides strong technical support for improving aviation safety and service quality.