This study proposes a fruit fly recognition method based on wingbeat sounds using autoencoder models deployed on a microcontroller. Spectrogram images were generated from audio data to train ANN-AE and CNN-AE models. Reconstruction error MSE and the Gamma distribution were used to determine the classification threshold. Before quantization, ANN-AE and CNN-AE achieved accuracies of 99.75% and 99.85%, respectively, while post-quantization accuracies remained at 99.60% and 99.65%. ANN-AE achieved a faster inference time of 1.3 ms per image. Although autoencoder models were more sensitive to noise, they showed strong potential for low-cost embedded intelligent insect traps.