Bachelor's Thesis

Speech/Music Classification

Author of thesis: Marek Hric

Acad. year: 2025/2026

Supervisor: Ing. Vladimír Malenovský, Ph.D.

Abstract:

This bachelor's thesis focuses on the problem of speech and music classification under real-time and causality constraints. The standard 2-class speech-vs-music formulation of the reference methods is extended with a third class for silence and ambient noise. The main objective is to design and implement a classifier satisfying these constraints and to objectively evaluate it against existing reference methods on a sufficiently diverse dataset. The first part of the thesis provides an overview of current approaches to audio signal classification, covering both traditional machine learning methods based on handcrafted features and modern neural network-based approaches. The second part focuses on the design and implementation of the dataset construction pipeline used to assemble the dataset for training and evaluation. The third part presents the experimental work. Four reference methods are implemented and compared using a benchmark on the constructed dataset. The subsequent experiments focus on the best-performing Temporal Convolutional Network (TCN). Based on these experiments, the proposed method called TCN-S is derived. The reference and proposed methods are then compared and analyzed. The results show that TCN-S achieves performance comparable to the reference TCN architecture while using only a fraction of the parameters and exhibiting lower computational latency.

Keywords:

speech and music classification, on-line audio classification, causal Temporal Convolutional Network (TCN), deep learning, machine learning, dataset construction

Date of defence

16.06.2026

Result of the defence

Defended (thesis was successfully defended)

znamkaAznamka

Grading

Process of defence

Student nejprve prezentoval výsledky, kterých dosáhl v rámci své práce. Komise se poté seznámila s hodnocením vedoucího a posudkem oponenta práce. Student následně odpověděl na otázku oponenta a na další otázky přítomných. Komise se na základě posudku oponenta, hodnocení vedoucího, přednesené prezentace a odpovědí studenta na položené otázky rozhodla práci hodnotit stupněm A.

Topics for thesis defence

For a three-class classification problem like in your thesis. What are the pros and cons of using three-binary classifiers versus a 3-way classifier? In other words, in your thesis, you have used 3-binary classifiers (sigmoid fn.) followed by an argmax over individual probabilities -- why not a 3-way classifier using a softmax function.
Kolik klasifikátorů máte?
Je dosažený výsledek prakticky užitečný?
Jak jste prováděl označování nahrávek?

Language of thesis

English

Faculty

Fakulta informačních technologií

Department

Department of Computer Graphics and Multimedia

Study programme

Information Technology (BIT)

Composition of Committee

doc. Ing. Lukáš Burget, Ph.D. (předseda)
doc. Mgr. Adam Rogalewicz, Ph.D. (místopředseda)
Ing. Libor Polčák, Ph.D. (člen)
Ing. Michal Hradiš, Ph.D. (člen)
Ing. Martin Žádník, Ph.D. (člen)

Supervisor’s report
Ing. Vladimír Malenovský, Ph.D.

The student demonstrated solid technical skills, particularly in dataset construction and experimental analysis, and proposed a lighter variant of a TCN model with competitive performance. The results are well supported by experiments, but the contribution remains mainly incremental. Overall, the work meets the task assignement, though it lacks some stronger originality.

Evaluation criteria	Verbal classification
Informace k zadání	This bachelor thesis addresses the problem of real-time speech/music classification. The overall difficulty is moderate, as the topic has been extenssively studied and there are numerous existing references in the literature describing baseline methods, evaluation metrics, and common approaches. The task combines the evaluation of classical methods and neural networks, implementation of the schosen approaches, dataset design, and experimental evaluation. The student worked independently, consulted regularly, and made a steady progress. A strong aspect of this work is the construction of a reasonably balanced dataset, together with the implementation and evaluation of several reference methods. The experimental part focuses on the Temporal Convolutional Network and leads to a simplified variant (TCN-S) with comparable accuracy and lower computational complexity. This represents a useful abd a practical contribution. However, the assignment also required proposing improvements over the state-of-the-art method, which was not fully achieved. The student explored only one neural architecture. The results are technically sound but mainly incremental. Despite the lack of originality, in my opinion, the thesis fullfills the assignment.
Práce s literaturou	The student worked well with the relevant sources on speech/music classification. Due to the lack of time he did not explore some of the latest state-of-the-art neural techniques in greater detail but this did not have an negative impact on the experimental part of his work. Based on his study, he was able to select and successfully implement the selected methods for evaluation. Overall, the student demonstrated the capability of studying and utilizing the available literature for an independent work.
Aktivita během řešení, konzultace, komunikace	The student was moderaltely active throughout the work. He regularly attended consultations, and was generally well prepared for them. He communicated his progress and responded well to my feedback. Teh agreed sub-tasks were mostly followed, and the student demonstrated a consistent effort during the development of the thesis. Overall, the level of activity, communication, and engagement was satisfactory.
Aktivita při dokončování	The thesis was completed very close to the submission deadline, and the final version was consulted only to a limited extent. While the main parts of the work were discussed during the development, the final form could have benefited from a more thorough consultation and refinement.
Publikační činnost, ocenění	To the best of my knowledge, there has been no publication activity or awards related to the thesis. The thesis is accompanied with an open-source software demo allowing thorough examination and possibly a follow-up.

Points proposed by supervisor: 77

Grade proposed by supervisor: C

Reviewer’s report
Santosh Kesiraju, Ph.D.

The work done in the thesis is very solid interms of the presentation and experiments. The presentation of the results is especially very clear and elegant.

Evaluation criteria	Verbal classification	Points
Náročnost zadání	Evaluation level: průměrně obtížné zadání
Prezentační úroveň technické zprávy	The report is fairly well written. All chapters are logically connected.	93
Formální úprava technické zprávy		95
Realizační výstup		95
Využitelnost výsledků	- The work might have practical usecases as it builds on existing works with an emphasis on practical usability -- real time classification while being computation budget friendly.
Rozsah splnění požadavků zadání	Evaluation level: zadání splněno
Rozsah technické zprávy	Evaluation level: je v obvyklém rozmezí
Práce s literaturou		95

Topics for thesis defence:

For a three-class classification problem like in your thesis. What are the pros and cons of using three-binary classifiers versus a 3-way classifier? In other words, in your thesis, you have used 3-binary classifiers (sigmoid fn.) followed by an argmax over individual probabilities -- why not a 3-way classifier using a softmax function.

Points proposed by reviewer: 94

Grade proposed by reviewer: A

Responsibility: Mgr. et Mgr. Hana Odstrčilová

VUT

Faculties and university institutes

Parts

Speech/Music Classification