Student literaturu získával samostatně a aktivně.

Aktivita při dokončování nebyla pod stresem. Text byl dodán ke kontrole včas. Vedoucím byly doporučeny dílčí úpravy struktury a jazyková kontrola.

Individual chapters of the work logically follow each other, and especially the sections presenting the data and results give the reader a good insight into the dataset and the basics of the machine learning approach to detecting dyslexia from speech. I would, however, appreciate a more detailed introduction to the problematic than the four small paragraphs here. Later in the text, I also have problems with some figures: 

• Fig 2.3 is taken from a paper where the depicted blocks make sense as they are explained. Here, the details are not given and the figure does not fit. It would require more details when explaining how the HuBERT model works so that the reader can understand the figure.
• When comparing systems on ROC or DET plots, the legend should be informative (corresponding to the caption). For example, Figs 4.5, 4.9, and 5.3 have almost identical text in the legend, not corresponding to the caption.

The work is certainly complete and functional. It indeed solves the assignment and is demonstrated online (and it was also demonstrated at the student conference). 

I appreciate the work with data, honesty with the discovered bias and cross-analysis with eye-tracking data.

I would appreciate, however, more attention to finding optimal approaches in the speech model and setting various baselines. In Fig. 2.2, we get a nice overview of what worked on detecting dysarthria, but I am missing some simple MFCC-based baseline, which could easily justify using models such as HuBERT here. Then, in every experiment, we get the result for features extracted from every layer of the HuBERT model, which is nice, but as it is common knowledge that these layers look at different kinds of speech characteristics, I would appreciate trying to leverage that, perhaps with some simple fusion (even as simple as a weighted average of features or even scores).

Also, I think that the author could compare SVM classifier to a simple neural network.

Evaluation level: zadání splněno

Overall, I am satisfied with the work and the approaches taken to analyze the dataset and the achieved results. One minor complaint could be perhaps the point where the student is asked to experiment with different architectures of neural networks, which we do not see much, but we certainly see various methods for detecting dyslexia.

The author correctly cites numerous sources throughout the work. In general, I have no issues with the bibliography in this work; I would sometimes recommend offering some of the details within the text where they matter instead of simply giving a list of citations. Like the already mentioned introduction.