Biodiversity loss is rampant worldwide, particularly in tropical regions like the Amazon. In the last decade, acoustic indices have been proposed as a rapid method for biodiversity assessment. However, their overall effectiveness as proxies for biodiversity is under debate. Here, we tested advanced statistical methods based on acoustic indices recently proposed to estimate species richness more accurately. Using an annotated audio dataset (2356 one-minute files) and song characterization of tropical bird assemblages from land-bridge islands in an Amazonian hydroelectric reservoir, we fitted both regression models and random forest algorithms to address the following questions: (1) do acoustic indices provide accurate estimates of bird species richness? (2) are univariate (a single index) or multivariate models (a combination of indices) better at predicting species richness? (3) at what temporal scale (minutes or hours) and with which measures (raw values or mean and standard deviation values)? (4) do these indices reflect spatial (island size) and temporal patterns (diel cycle) of singing activity? Although we found that multivariate models using a set of acoustic indices computed at a broader scale (hours) performed better than simpler models, their overall predictive power of species richness was poor for these tropical bird assemblages. The high heterogeneity and variation in the acoustic activity and signals of the Amazonian bird species present a considerable challenge for acoustic indices to capture changes in species diversity adequately. In agreement with recent studies, our findings point out the limits of acoustic indices, especially in tropical, highly diverse regions, emphasizing that caution should be used when applying this type of acoustic indices in biodiversity assessment. In contrast, all tested indices reflected distinct spatial and temporal patterns that were often related to habitat features (i.e. island size) and animal activity (i.e. choruses), supporting alternative (large-scale) applications of acoustic indices. Random forest algorithms confirmed the potential to classify island size based on soundscape characteristics. These findings suggest acoustic indices can capture differences in assemblage composition and bird activity along a habitat fragmentation gradient. Hence, they can more efficiently assess habitat and community structure than species diversity, occurrence, or abundance.