When operating a wind turbine, damage of rotor blades is a serious problem. Undetected damages are likely to increase overtime, and therefore, the safety risks and economical burdens also increase. A monitoring system, which detects reliably defects in early stages, gives scope for action and is therefore a key element to avoid damage increase and to optimize the efficiency of wind turbines. One promising approach for damage detection is acoustic emission methods. Although most acoustic emission approaches use ultrasonic sound waves of the structure that require about 12 to 40 sensors to monitor one rotor blade, we propose to use the airborne sound in lower frequencies from about 500 Hz to 35 Hz with three optical microphones and present a signal model-based damage detection algorithm. The real-time algorithm uses six audio features from a spectrogram representation to detect damages and to estimate its significance. In a fatigue test of a 34-m blade, the algorithm detected the damage event and damage increasing without false detection. It was also tested with recordings inside an operating blade of a 3.4-MW wind turbine. In the recorded time period of about 1 year, the algorithm indicated no false detection. © 2020 The Authors. Structural Control and Health Monitoring published by John Wiley & Sons Ltd