1. Introduction to the Thesis
András Dörömbözi wrote “Audio Tampering Detection: Deep Learning methodologies for multi-layered threat detection” as his diploma thesis in Data Science at TU Wien and handed it in July 2023.
The thesis asks a simple but important question. How can we tell that a speech recording has been edited, for example by cutting out a sentence? Many existing methods rely on the faint hum of the power grid that gets recorded along with the voice. The author points out that this stops working in everyday situations, such as recordings made on a phone, or when music is laid over the speech, as is common on TikTok. So he tests how well the classic method copes with such recordings, and whether neural networks can do better.
The thesis follows a clear order, starting with an overview of existing methods, the setup of the experiments, the results, and a discussion that answers three research questions.
The practical part is not a finished app. Instead, the author created his own test recordings (speech with added music, background noise and smooth fade-outs at the edit), rebuilt the classic method from a research paper, and trained several neural networks. The trained models were later used in a research project called defalsif-ai. I could only judge this from the text, not from the code.
In short, the classic method works very well on clean recordings but struggles once music or noise is added. The neural networks were more stable in some cases, but none of them was clearly better overall.
2. Evaluation of the Academic Quality of the Thesis
Quality of design and execution
The idea is clear and sensible. The author sets research questions, picks a well-known method as a reference, and checks that his own version of it gives the same results as the original paper. What I found less convincing is the size of some test sets. With so few recordings, a couple of lucky or unlucky guesses can change the picture. Some networks also seem to have learned to spot the fade effect rather than the edit itself. To his credit, the author notices and admits this.
Degree of innovation
The networks themselves are not new, since they come from other audio tasks. What is new is the problem, meaning edits that are covered up by added music or noise, plus the test data the author had to create for it. That makes the topic feel fresh and realistic.
Independence
It looks like a lot of independent work. He rebuilt the reference method, made his own datasets and tried many different models. Since it was part of a larger research project, I can’t tell exactly how much of the direction was his own idea, but nothing suggests he needed much hand-holding.
Organisation and structure
Easy to follow overall, since the questions come first and are answered in the discussion. The overview of existing work is quite long, and the results chapter is a long series of tables, so I had to search for the main message.
Quality of communication
The text is understandable, but the sentences are sometimes long. The discussion chapter is the clearest and easiest readable part.
Scope of the thesis
The scope is big for a diploma thesis, with two sets of recordings, several types of manipulation and many model variations. Maybe a bit too much, because a few surprising results are left unexplained.
Spelling, care, and accuracy
The thesis is put together with care. Tables and figures are numbered and explained, and the results are reported openly, even the disappointing ones. The few small slips, like an odd label in a table or a typo, never get in the way of understanding.
Use of literature
The 69 sources cover the main approaches well. A few basic explanations come from websites instead of textbooks, which I would have avoided.
3. Final Conclusion
Overall, this is a solid and honest thesis on a topic that matters. Its strongest contribution is showing that a well-known method can fail under realistic conditions. The results are mixed and the presentation could be more careful, but the author is open about both.
4. Relevance to My Own Final Thesis
I do not yet have a specific topic for my own final thesis, but I would like to move in the direction of audio forensics in the context of generative audio, looking at either acoustic environments or voices. This thesis is a useful starting point for that plan.
It shows that classic forensic traces, such as mains hum, room signature and microphone fingerprint, break down once recordings are layered or post-processed. Generated audio raises the same question in a sharper form, since a synthetic voice or an artificial soundscape may lack these traces entirely, or imitate them convincingly. Testing which of them still hold up on synthetic material is a promising open question.
The thesis also offers two practical lessons. Detectors can learn shortcuts, like the crossfade, instead of the real signal, so datasets and splits need careful design. And because no suitable data existed, the author synthesised his own, which suggests I will probably have to build datasets for generative audio myself. His pipeline gives me a template to start from.
