Sleep apnea syndrome (SAS) affects an estimated 3–7% of the global population, yet remains frequently undiagnosed. It manifests as interruptions in breathing lasting at least 10 s during sleep, caused by partial or complete obstruction of the airways. The current standard diagnostic method for SAS is polysomnography (PSG), an invasive procedure that relies on subjective assessments by clinicians. To address the shortcomings of PSG, our solution proposes a decision support system utilizing a tracheal microphone for data collection. We employ a deep learning (DL) approach named S-CRNN, which integrates a convolutional neural network (CNN) framework with a bidirectional Gated Recurrent Unit (GRU). This system analyzes log-Mel audio spectrograms and is trained under the siamese paradigm. Final detection of apnea events utilizes an unsupervised clustering algorithm, specifically k-means, applied to S-CRNN-processed data. Validation with data from eight patients yielded a \(\textit{Recall}\) rate of 90%, a \(\textit{Precision}\) of 88.4%, and an \(\textit{F1-score}\) of 89.2%. Comparative analysis against contemporary methods underscores the efficacy of our siamese training approach in supporting SAS identification.