The thesis aimed to develop an extension for existing music/vocal separation tools to support subsequent speech transcription. The thesis is well structured and follows a clear, logical progression. The student clearly describes the tools used and outlines their principles of operation.

The student's approach to the thesis workflow was highly problematic. Despite the supervisor's continuous efforts to establish regular cooperation, not a single meeting took place during February, March, and April. The student worked on the project almost exclusively during the last 5 weeks before the submission deadline. While a few consultations did eventually take place and the student incorporated some of the supervisor's feedback, I did not have the opportunity to see the final version of the thesis before its submission.

Furthermore, the implementation shows signs of heavy reliance on AI tools, which has unfortunately resulted in a weaker understanding of the generated code. The resulting tool is functional, though it has visible reserves. Due to the lack of continuous consultation, it is highly appropriate and necessary for the student to thoroughly present and demonstrate the final version of the application during the defense.

There is almost the entire source code listed directly in the main text, moreover without syntax highlighting. Standard academic practice is to place full source code in an electronic appendix or public repository, while the main text should include only essential snippets, supported by flowcharts or pseudocode to explain the algorithms.

Questions:
1. What explains the low number of “successful runs” and the relatively high number of “failed runs” reported in Chapter 6?

2. What does the WER (Word Error Rate) metric represent, and how do you explain the low values reported in Chapter 6.4?

3. Is the resulting tool publicly available (e.g., in a public repository)? Points proposed by supervisor: 77

The aim of the bachelor's thesis is to create software for music and vocal separation with a focus on French rap for subsequent automatic transcription. Student implemented this software as a web application using third-party tools, including three different separation tools. The Whisper tool is used for subsequent transcription. Verification was performed in a manner of evaluation of transcribed text using the WER metric using a third-party tool.
I do not consider the use of tools to be a problem, the burden of the work lies in the implementation and testing of the resulting application. However, the separation methods and metrics should be discussed in more detail in the theoretical part. The thesis presents only a superficial principle of the methods illustrated by adopted images without their description.
I consider the weak point of the work to be the testing, where the success rate of the transcription is not separated from the quality of the vocal separation. Although it is probably difficult to find a recorded track of French rap, it is necessary to test the success rate of the Whisper tool when using this type of input audio data, or compare it with English rap, in order to show the influence of language and slang on the success rate of the transcription. Furthermore, the separation itself should be evaluated using objective metrics, e.g. SDR or similar ones, where the ground truth would be a track with rap. It should be compared with the separated vocal, where the input would be the same track with rap mixed with musical accompaniment. A subjective evaluation of the separation would also be beneficial, e.g. using C-V-C tests or Diagnostic Rhyme Test. It is a pity that the audio outputs of the separation are not included in the appendix.
The work has a clear structure and is logically structured. However, in my opinion, chapter 1 should be part of chapter 3 and the text should start with the theory of vocal separation. The images used are properly cited, but they should be redrawn into vector graphics. The text in the images should be the same size as the text of the paragraphs. The sources are properly cited, however, some statements in the text lack references to literature. Images and code listings should have been better positioned in text to avoid the blank spaces and unnecessary division of code listings at the end of the page. The codes in the listings should be commented. Topics for thesis defence:

1. What objective metrics besides SDR would you use to evaluate the quality of separation of the vocal part? As part of your presentation, present audio samples to compare methods of separation of the vocal part.

Points proposed by reviewer: 85