voice biometrics – Biometriq

The exhibition was marked by the ubiquitous AI. Many companies presented their latest achievements in constructing systems that communicate autonomously with people. The humanoid robot Ameca (Etisalat) interacting with its interlocutors aroused great interest. The stands with interactive agents (Amdocs) offered an almost unbelievable quality of image and speech generated by the systems.

Google has unveiled Gemini Live, its response to ChatGPT’s voice mode. Gemini Live has function Share Screen With Live, that allows Gemini to interact with the image displayed on the phone’s screen. Deutsche Telekom has indicated a possible direction for the development of phones by turning the entire phone into a chatbot. The phone has no applications and is a personal assistant that communicates with the user by voice. The basis of the solution is a digital assistant from AI Perplexity, but it is also to be open to, among others, Google Cloud AI, ElevenLabs, and Picsart. South Korean startup Newnal has presented a new operating system for mobile phones that uses historical and current user data to create a personalized AI assistant that is to eventually become an AI avatar behaving just like the user.

All of the above solutions, as well as many others, are connected by the use of voice technologies for two-way communication. The direction indicated at MWC 2025 is clear – our actions will be supported by avatars and bots communicating with us autonomously. The possibility of quick, machine confirmation of who we are talking to is therefore becoming even more important than ever before, because the quality of autonomous voice communication systems does not guarantee correct verification of the speaker by a human.

Photos by Andrzej Tymecki

How to effectively detect voice-based fraud? How to distinguish a real voice from a fake one, e.g. one generated on the basis of AI? The answer is simple. This requires advanced voice biometrics tools and a number of analyses. We publish here two examples that we analyzed some time ago in our laboratory and, thanks to our proprietary algorithm, we assessed with a very high probability whether the voice is real or fake and to what extent it is consistent with the voice of a given person.

The analyzes concern:

● recognizing the voice of one of President Duda’s Russian pranksters pretending to be President Macron

● assessment of the similarity of the voices of actors Piotr Fronczewski and Filip Pławiak in the film Rojst. In the play, men play the role of the same person (Kociołek) in adulthood and youth, respectively.

We share with you the conclusions from these experiments.

Biometric comparison of the voices of Fronczewski and Pławiak.

For this purpose, we used 25 seconds of total speeches by both characters, composed of several fragments of their original speech, based on the original film soundtrack. What compliance did we achieve?
The results of the analysis showed that the actors’ biometric voices are NOT consistent. Pławiak statement vs. Fronczewski VP – only 15% agreement, Fronczewski statement vs. Pławiak VP – 11%, but interestingly, these differences are not noticed at the level of the ear. In our opinion, the voices of Pławiak and Fronczewski are almost identical. And that is ultimately what this is all about.

For both characters, gender and nationality were recognized with minimal uncertainty (score of almost 100%). An age difference between the characters was also detected, estimated at 20 years.

Analysis of the voices of Russian pranksters Vladimir Kuznetsov (Vovan) and Alexei Stolyarov (Lexus) impersonating President Macron.

In this case, we biometrically analyzed the recordings of the pranksters’ voices and compared them with the voice of the real Macron (in both Polish and English versions). We downloaded all voice samples in the form of individual recordings from the public domain on YouTube. Our goal was to confirm the effectiveness of biometric systems for this specific situation – identifying fraud.

It turned out that the voice of one of the “Lexus” pranksters was just over 50% consistent with the voice of the President of France and as much as 97% consistent with the voice of the false president. The voice of the second one – “Vovana” – showed no similarities (0%) to the fake president.

This clearly proves that thanks to biometric analysis we managed to:

●detect the fact, only after 1 minute, that a fake president was involved in the conversation
● identify the identity of the fictional president (Lexus)
● confirm that the public domain is a very good source of voice samples, which may not always be used for noble purposes
● strengthen the thesis that the most effective attacks are those using social engineering, and in this case it was the choice of the right time when the President was faced with increased stress (rocket fall).

These are just selected examples of the use of specialized biometric tools to confirm the identity of people. If implemented in the future, they may help detect voice-based abuse.