Basecalling, similar to many other domains of machine learning, suffers from the problem of having to accept a trade-off between prediction speed and accuracy.
Bonito, based on the architecture of QuartzNet, shows similar results to Guppy, which is considered to be state of the art for basecallers.
The convolutional architecture of Bonito, however, has the potential to reduce the prediction time markedly compared to the recurrent architecture of Guppy.

This work attempts to provide insight into the effect of tuning the hyperparameters available in Bonito.
This effort is made with the focus of improving the speed of predictions without negative impact on the accuracy.
In order to alleviate the problem of reduced accuracy in smaller networks, we apply knowledge distillation, which, in other domains, is shown to improve accuracy.

The results of our experiments suggest that dilation, combined with a reduced kernel size, can improve prediction speed and accuracy of Bonito.
Additionally, we show that knowledge distillation can improve the accuracy of basecallers.
Notably, the most significant improvements are observed on large basecallers.
Nevertheless, the results suggest that knowledge distillation should always be applied for any size of basecaller.