Izabela Paniczek won the third place in the CreatiWITy 2026 competition for her project on detecting manipulation in voice recordings. The system uses artificial intelligence methods and supports the user in analysing the authenticity of audio recordings.

The project “Development of a system for analyzing the authenticity of voice recordings” was carried out under the supervision of Prof. Krzysztof Brzostowski. Its aim was to develop a user-friendly application for the detection of manipulations in audio recordings.

Izabela, a graduate from System Engineering, explains that it was important for her to address problems related to sound. Previously, while working at the Neuron student research club, she was involved in generative music models.

The growing problem

The inspiration to address the topic came from the increasing scale of sound manipulation both in public spaces and in everyday life.

“We increasingly often encounter recordings with fake content. With the help of AI, we can very closely imitate or generate human voice,” says Izabela Paniczek.

She emphasizes that the issue affects not only well-known individuals and politicians, but also ordinary people. Computer-generated voice is also a growing problem in the case of telephone conversations, in which the mechanism can be used in fraud.

“A certain number of voice samples is enough to reproduce it. In the case of public figures, this is even easier because there is greater access to their voices. Such threats also start to affect everyday life,” adds Izabela.

The increasing availability of voice-generating tools means that producing realistic recordings no longer requires specialist knowledge. Voice imitation depends on the available computing power. The better the processor or graphics card, the greater the potential for achieving natural voice.

“Text-To-Speech models are widely available today, often even for free. Although the sound can still be slightly robotic, in practice many people are unable to distinguish it from a real recording,” she emphasizes.

Users lack tools

Although sound generation and modification technologies are developing dynamically, solutions that allow the detection of manipulations are not widely available.

“I have noticed that there are many text analysis tools that attempt to detect AI-generated content. When it comes to sound, such widely available solutions are not available,” explains our student.

The available systems are mostly commercial solutions aimed at businesses. There is a lack of applications for individual users.

Record-analysis system

Izabela Paniczek has addressed this problem by developing a system that analyses the authenticity of voice recordings. The tool is available in both mobile and desktop application forms.

The user can upload an audio file or record sound directly in the app. It then analyses the recording and determines the likelihood of its authenticity. The model does not provide only a simple “yes” or “no” answer – it indicates the percentage probability that the recording is authentic or manipulated.

The model was trained, among others, on popular datasets used in scientific research, for example on ASVspoof 2019 and In-the-Wild. The dataset was expended by implementing such techniques as changing pitch and adding noise, so that the model is better prepared for previously unknown voice manipulation methods.

The author tested the model on both publicly available datasets and her own recordings, which also included examples of imitating the voices of well-known individuals. A total of over ten thousand audio samples were used.

Decision explainability

The analysis result is presented, among others, in the form of a heat map on the spectrogram, indicating the parts of the recording and frequency ranges that had the greatest impact on the assessment. Additionally, the system generates a textual description with an interpretation of the analysis results.

“The user can view which elements of the recording influenced the system's decision and go back to particular fragments of the recording. In effect, the tool not only provides analysis results but also ensures that they can be trusted,” explains Izabela Paniczek.

She emphasizes that the because of the development of generative technologies, manipulation-detection systems need to be constantly updated.

“This is not a solution that can be developed once and for all. Voice-generating models will be better and better, which is why regular updating and training on new data is necessary,” she points out.

Currently, the student's application is doing well in detecting manipulations, however, further work is planned on its development and improvement of some of its elements.