Results of the mixed model.

In recent years, several technologies have been developed for the monitoring and control of insect vector species. Many of them aim to use mosquito wingbeat frequency in the form of sound or opto-acoustic measurements to identify mosquito species, often through the training of AI classification models. However, these models often struggle to be accurate in real-life conditions, as the training data rarely captures the variability range of different species across many individual and environmental conditions, or does not explicitly control for it. Here, we use lab recordings of mosquito sounds to evaluate the impact of several environmental and life history factors on the mean frequency of the first harmonic of mosquito sounds. We recorded 475 individuals of 15 species in several environmental conditions, varying in temperature and humidity, while we also characterized the effect of body size (wing length), sex and age on the frequency of wingbeat sound at the among-individual level. Only species that comprised at least 2 recorded individuals were included in the analysis (N = 10 species). Variances at the within-individual and within-species level varied consistently, as the repeatability of the trait was 0.411 and 0.466, respectively. However, when we controlled for morphological and environmental effects, the proportion of between-individual variance decreased, while the between-species component increased (repeatabilities: 0.267 and 0.630). This suggests that species-specific signals in the sound are more robust once factors introducing variances due to real life conditions are involved in the models. Sex and temperature both had a significant effect on mosquito sound: an increase in temperature led to an increase in wingbeat frequency. In addition, the random slope analysis showed that response to temperature differ between species, with strong between-species differences, especially for males. Therefore, advancing AI species recognition requires that biotic and environmental variables be either explicitly integrated into classification models or sufficiently represented in training data to reflect real-life variability.