Automatic detection and identification of pharyngolaryngeal activities (PLA) such as swallowing, speaking, and coughing, are crucial steps in developing a system for diagnosing dysphagia (swallowing disorders) using high-resolution cervical auscultation (HRCA) signals. Using sensors placed on the throat, HRCA provides clinicians with a non-invasive means of assessing swallowing. This study introduces a new approach based on depthwise CNN and LSTM networks to automatically detect and identify PLA in synchronized accelerometer and microphone signals. This method achieves an Fscore of 88% for swallow detection across seven hours of HRCA signals recorded from 42 healthy subjects under both supervised and ecological conditions. It outperforms state-of-the-art systems in real-world settings. On other activities, the system reaches an Fscore of 93% for phonation and 86% for airway defense mechanisms, making it a promising tool to assist clinicians in the analysis of HRCA signals and a reliable base towards non-invasive assessment of dysphagia.