MiniMax Surpasses Google: Voice AI Deceives 66% of Listeners

MiniMax Challenges Tech Giants with Its Voice Synthesis

In a world where voice synthesis technology is rapidly expanding, a Chinese startup named MiniMax has managed to stand out spectacularly. This company, still relatively unknown in Europe, has outpaced giants like Microsoft, Google, and Amazon in the field of artificial voice creation. This achievement comes from a recent study that revealed two-thirds of listeners failed to identify the AI-generated voices from MiniMax as artificial.

Voice synthesis is a growing sector, often dominated by large tech companies that have heavily invested in the development of this technology. However, a study conducted by Vocal Image, a startup based in Estonia, has highlighted an unexpected reshuffling of the cards in this domain.

MiniMax Tops a Comparative Study

The study involved 20 voice synthesis models, tested by over 10,000 participants in the UK and the US over a month-long period. The listeners, unaware that they were listening to AI-generated voices, were asked to evaluate each voice based on 18 different criteria, such as warmth, clarity, monotony, and perceived confidence.

MiniMax, supported by Chinese giants like Alibaba and Tencent, saw its Speech-02 model rise to the top of the rankings. This model was deemed the most realistic and credible, surpassing those from Google, Amazon, and Microsoft. Notably, 86% of native English speakers placed Speech-02 in first place. This result is particularly impressive given that the British, according to the study, are the most adept at detecting artificial voices, with a 13% higher effectiveness than Americans.

The quality gap between the different models tested is also significant. The highest-rated model achieved a score three times that of the lowest, highlighting a notable disparity in the quality of voice syntheses.

The Implications of Human Perception

The fact that 66% of listeners did not detect the artificial nature of MiniMax's voices is not merely a technical feat. It sheds light on a potential flaw in human perception, which could have significant security implications. The study showed that when listeners identify a voice as artificial, they tend to reject it. Thus, the ability to perfectly imitate a human voice becomes a crucial issue.

Nick Lahoika, founder of Vocal Image, emphasizes that large tech companies have failed to master this aspect. Their models, designed for mass and general use, lack the finesse necessary for contexts requiring precision and emotional nuances. Lahoika refers to the "last mile" of voice quality, which includes elements like number pronunciation and contextual intonations, often overlooked in favor of scale.

A Threat to Digital Security

Beyond the listening comfort, this technological advancement raises security concerns. Reports of number spoofing on the "J'alerte l'Arcep" platform have skyrocketed, jumping from 531 in 2023 to over 19,000 in 2025. Fraudsters use numbers from banks or public authorities to make their calls more credible, a technique made more effective by indistinguishable synthetic voices.

In response to this growing threat, Arcep decided to open an administrative investigation on January 29, 2026. This investigation aims to verify whether operators are complying with caller ID authentication obligations. The ability of a startup to produce voices undetectable by the human ear, accessible via API and at low cost, is not just a technical achievement. It is also a potentially dangerous tool in the hands of fraudsters.