Photovoltaic installations are facing various maintenance challenges nowadays, particularly with regard to real-time monitoring and anomaly assessment. Consequently, researchers have explored and improved several traditional approaches with different levels of efficiency. In this study, however, the main focus is placed on presenting a potential non-conventional, sound-based method intended to improve overall reliability. Precisely, this manuscript introduces a first experimental investigation demonstrating that acoustic signals may serve as an accurate, non-contact indicator for certain hazardous breakdowns in photovoltaic inverters or any other installation involving electrical components. To this end, a data set, comprising 2,000 waveform samples of abnormal conditions, each with a 3-s duration, and an additional 2,000 samples taken under normal operating conditions, was carefully collected. Then a basic Convolutional Neural Network (CNN) model was designed, with its final dense layer returning the probability of an individual sample falling into either the normal or abnormal class. The final results and metrics were promising with a precision achieving 0.9802, and a loss equal to 0.0505, providing satisfactory evidence that monitoring applications based solely on acoustic stream could be developed and incorporated within intelligent control frameworks for early electrical hazards detection in photovoltaic systems.