I am pleased that Karel Beneš has submitted his thesis under my supervision, culminating his years of effort into the submission of his doctoral dissertation at FIT BUT.

Karel’s thesis addresses the important problem of utilizing language models to support recognition systems, such as those used for automatic speech recognition and optical character recognition. More specifically, the work focuses on scenarios where the language model is trained on imperfect or erroneous training data, or is applied to postprocess the erroneous outputs of the recognition system. It analyzes how these errors influence the behavior and effectiveness of language models and proposes methods to mitigate their impact. The scientific content of the thesis is well-articulated in the document and its reviews; therefore, I will concentrate on more personal remarks.

Karel is an excellent student (he graduated both the Bachelor and Master degrees at FIT with the highest distinctions — the “red diploma”) and I am happy that he has been part of our laboratory since his Bachelor studies; his thesis “Finite State Grammars and Language Models for Automatic Speech Recognition” supervised by Dr. Mirko Hannemann (currently with Apple, Inc.) dates back to 2014. He has been actively working in the area of neural artificial intelligence models, applicable across various fields including automatic speech recognition (ASR), natural language processing (NLP), and optical character recognition (OCR).

Karel has authored 15 conference and journal publications, most of which have appeared in respected international venues such as Interspeech (a CORE A conference) and the International Journal on Document Analysis and Recognition (IJDAR), a leading journal in automatic document processing. His research has been widely cited; according to Google Scholar, he has received 53 citations, a commendable achievement for a researcher at this stage in his career. His work on the “Residual Memory Network” earned him and Murali Karthick Baskar the Best Student Paper Award at INTERSPEECH 2017 in Stockholm.

His contributions are also prominent in various projects. Karel was a valuable team member of the Neural Representations in Multi-modal and Multi-lingual Modeling (NEUREM3) project, funded by the Czech National Science Foundation (GACR) under the prestigious EXPRO scheme. Additionally, he participated in the EU Horizon 2020 project, Multiple Intelligent Conversation Agent Services for Reception, Management and Integration of Third Country Nationals (WELCOME), where he applied his expertise in speech and language technologies to handle dialogues in under-represented languages. His involvement in a series of OCR projects sponsored by the Czech Ministry of Culture stands out; the resulting application integrates advanced computer vision and NLP techniques, enabling automatic transcription of various printed documents in most European languages, including Latin, old documents in Fraktur and similar scripts in German, and handwritten Czech. This application provides an efficient interface for text corrections and offers multiple transcription formats, and it is routinely used by several Czech and European libraries, including the Military History Institute Prague and the University Library of Mannheim.

Karel is also actively looking to enrich his knowledge beyond his native lab, and he was with RWTH Aachen (the most respected German speech and NLP laboratory) for 6 months in 2018/19 working on two-pass decoding in automatic speech recognition with Kazuki Irie (now at Harvard University).

Moreover, Karel is an excellent and dedicated teacher, coordinating the compulsory Artificial Intelligence and Machine Learning (SUI) course for all Master’s students at FIT BUT. Despite being a Ph.D. student, he ranked 3rd in the 2023 FIT teacher rankings for the Master’s program. He never misses an opportunity to help students (and sometimes also seniors) in the lab on a variety of issues ranging from Python programming, through running experiments on massively parallel computing architectures, to medieval swordsmanship. He is also a keen organizer of the group’s sports life including its indoor climbing club.

In conclusion, I wholeheartedly recommend Karel Beneš’s Ph.D. thesis for defense. I wish him all the best in his future professional and personal endeavors and look forward to continuing our collaboration.

In summary, the thesis integrates novel approaches, real-world applications, and high scientific rigor to address key challenges in language modeling. While there are areas for improvement in formal presentation and broader applicability, the work meets the criteria for awarding a doctoral degree. The publications and research outputs further validate the candidate’s contributions to the field. In my opinion, the thesis and the student's achievements until now meet the generally accepted requirements for the award of an academic degree. Topics for thesis defence:

1. How were the hyperparameters for NN training selected? E.g. LR = 20 for LSTM (Chapter 3.5) seems unusually large. Was any sweep of the hyperparameters performed?
2. „Note that this way we never use the introductory paragraph for i-vector estimation. This way, we make the task more challenging“ (Chapter 3.6) It seems intuitively correct, but has it been tested?
3. Has there been a consideration to use the Transformer model as a LM? Could you extend the solution to them? Why YES/NO?
4. Softmax is never defined - please define and explain.
5. Discussion. There are several methods presented in this work. Can the candidate present one as the main outcome of his work? Can you elaborate on the answer?

Karel Beneš’s publication record, consisting of 15 papers, awards, including a best student paper at Interspeech, his teaching excellence, and experience participating in international challenges such as CHiME, IWSLT, etc., demonstrate that he meets accepted requirements for being awarded a Ph.D.