Fetal Breathing Movements (FBM) serve as a critical indicator of fetal neurological development and intrauterine health. Despite their clinical importance, existing FBM monitoring techniques such as real-time ultrasonography and invasive tracheal pressure sensing remain limited by high operator dependency, low temporal resolution, and poor suitability for continuous or early-stage monitoring. To overcome these limitations, we propose a novel deep learning framework, Attention-Weighted Spectral Token Classification (AWSTC), designed for accurate, non-invasive detection of FBM using fetal Doppler acoustic signals. The proposed method incorporates a high-resolution preprocessing pipeline involving Wiener filtering, amplitude normalization, and Short-Time Fourier Transform (STFT) to generate robust time-frequency spectrograms. These are partitioned into spectral-temporal patches and encoded with positional embeddings, allowing the model to preserve fine-grained temporal dynamics and spectral locality. A multi-head self-attention mechanism enables the model to capture long-range dependencies, while a classification token aggregates global context to infer FBM presence. AWSTC outperforms baseline models, CNN-LSTM (90.2%, 0.920), Random Forest (89.3%, 0.880), and SVM (88.1%, 0.850), achieving state-of-the-art performance with 93.6% accuracy and an F1-score of 0.965. In addition, attention-based saliency maps highlight physiologically relevant spectral regions, enabling interpretability and clinical transparency. The proposed framework offers a scalable, explainable, and real-time solution for fetal respiratory monitoring, advancing the integration of AI into obstetric care and establishing a foundation for next-generation, non-invasive prenatal diagnostics.