During out-of-hospital cardiac arrest, it is difficult to understand the effectiveness of cardiopulmonary resuscitation (CPR) and to know the condition of the patient without having had medical training and the use of specialist equipment. As a result, rescuers may non-invasively check the person's pulse using their hand, which can be unreliable. Our objective was to assess the potential of recording and using thoracic audio to automatically identify resuscitation events. Twelve (12) digital thoracic stethoscope recordings from a porcine cardiac arrest simulation study were reviewed and annotated according to simultaneous physiological signals. Each recording was annotated for at least one occurrence of normal sinus rhythm (NSR), ventricular fibrillation (VF), CPR and defibrillation. A 5-second audio epoch was extracted for each annotation. The 28 kHz audio data was compressed using a discrete wavelet transform with the Daubechies 4 wavelet function to 3.5 kHz. The time and frequency composition of the exported epochs were characterised. During NSR two distinct sound pulses were identifiable corresponding to S1 and S2 activity in the 10 to 100 Hz frequency range with a time-period of 0.2 s. Upon VF induction, the audio signal displays no distinct physiological features with the exception of mechanical ventilation used to support anaesthesia. The inspiratory phase of ventilation was identifiable at 150 Hz and prolonged for approximately 1 s. The application of CPR resulted in sound signatures in the frequency range of 10 to 1000 Hz and were sustained for approximately 0.3s; however, motion caused by CPR disturbed the contact between the stethoscope and the skin of the test subjects. As a result, events occurring after the application of CPR were not distinctive enough for characterisation, such as defibrillation outcome. Thoracic audio shows potential in classifying events during resuscitation and may provide a non-invasive method for detecting the return of spontaneous circulation.