INTRODUCTION Current methods for evaluating dysphagia have limitations due to their cost, invasiveness, or inaccessibility. High-Resolution Cervical Auscultation (HRCA) is a promising approach that aims to overcome these constraints. This method uses non-invasive sensors on the patient's neck and analyses the captured signals to obtain useful diagnostic information. However, analysing large amounts of signal data, particularly during ecologically relevant situations such as meals, can be challenging and time consuming for medical practitioners. In this study, we introduce a new algorithm that performs the automatic analysis of HRCA signals to identify and locate three specific pharyngolaryngeal activities (PLA): swallowing, phonation and airway clearance mechanisms (cough and throat clearing). MATERIALS AND METHODS The study involved 42 healthy subjects from 21 to 57 years old (mean 34 y. o.) who were recorded using the HRCA medical device Swallis DSA. This device provides sound and vibrations from the throat using a microphone and a tri-axis accelerometer. Participants performed the same supervised tasks including food intakes of five different textures (saliva and IDDSI 0, 3, 4, 7), coughing, throat clearing, and saying a sentence. Additionally, many spontaneous PLA were recorded between tasks. The algorithm is divided in two steps: an activity detector that removes silent or low-energy segments from the signals and a classifier to identify the PLA in the remaining segments. The classifier, which combines convolutional and recurrent neural networks, takes spectrograms of the four sources (microphone and the three axes of the accelerometer) as input and classify each activity segment as swallowing, phonation, expectoration or irrelevant. A cross-validation was conducted on the entire dataset to evaluate the algorithm's performance. RESULTS The proposed method achieved good activity segment classification score of 87.4% with F-score of 84.7% for swallowing, 86.8% for phonation, and 84.1% for airway clearance mechanism on the global 5.4 hours corpus containing 1704 swallows, 128 coughs and throat clearings and 1266 phonation events. The average recall for the three classes reaches 93.7%. Note that 92.6% of supervised swallows were detected and 85% of the spontaneous ones. DISCUSSION This method allows obtaining the location of events of interest and their sequence in long recordings. The actual behaviour of this algorithm tends to correctly detect most of the PLA but is overpredicting some of the events, particularly during periods of intense synchronised vibration and sound, such as chewing. This aspect could be targeted using specific filter processing. The combination of the two sensors for event detection leads to stronger robustness to background noise as no confusion appeared between phonation from the operators and the participant. This is an essential aspect to use this method in ecological conditions. The use of this non-invasive device with this algorithm could instantly provide medical practitioners useful information such as the occurrence of coughing after swallowing, the frequency of natural swallowing, and other key indicators in the assessment of dysphagia.