The invention discloses an audio howling suppression method, device and system and a neural network training method, and the method comprises the steps: obtaining a plurality of continuously collectedfirst audio frames in a time domain; converting the plurality of first audio frames into a plurality of second audio frames respectively, wherein the second audio frames are audio signals in a frequency domain, and each of the plurality of second audio frames comprises a plurality of frequency points; determining howling frequency points with howling signals in the plurality of second audio frames, wherein the howling signals are continuous, single-frequency and high-energy signals; in the second audio frame with the howling signal, removing the audio signal of the howling frequency point; inputting the frequency point signal near the howling frequency point in the plurality of second audio frames into the howling frequency point restoration neural network model, and outputting the restored howling frequency point replacement audio signal to replace the audio signal at the howling frequency point position. Howling signals are suppressed relatively easily and simply, fault of audio signals is avoided, and distortion of the audio signals is reduced.