OpenAI has announced a series of voice models called GPT-Live. These models have the ability to perceive the user's speech and generate a response at the same moment, making voice interaction with the neural network much more natural. The model can now interject its remarks while the user is speaking, adapt to their pace and pauses, and correctly react if interrupted.
The first implementation of ChatGPT's voice mode, released in 2023, functioned as a sequence of three separate components. First, the user's voice was converted into text using a speech recognition model, then a large language model created a text response, which was voiced by a third model. Although this cascade allowed for the integration of the most advanced language models from OpenAI at that time into a speech interface, the dialogue suffered from pauses and loss of intonation, which reduced its naturalness.
In the Advanced Voice version, introduced in 2024, the process of sound analysis, processing, and generating a voice response was integrated into a single multimodal neural network. This significantly reduced pauses and increased speech naturalness. However, even in this mode, communication remained strictly turn-based: the model waited for the user to finish their phrase before starting to speak. This led to problems when background noise or pauses in speech could confuse the model, causing it to switch between listening or speaking modes, disrupting the flow of communication.
The new GPT-Live model was introduced on July 8, 2026, and is equipped with a full-duplex architecture designed to eliminate the problem of rigid dialogue sequencing. The system no longer breaks the conversation down into separate turns; instead, the incoming audio stream is processed by the model in parallel with the formation of its own response. The neural network analyzes the stream multiple times per second, making decisions about actions: speaking, remaining silent, ignoring the interlocutor's pauses, or reacting to an interruption of its own speech. While listening, the model can use interjections such as 'uh-huh' or 'I see,' creating a feeling of complete attention to the user.
Thanks to the full-duplex structure, GPT-Live is capable of real-time synchronous translation between languages and better isolating the user's voice from external noise. During a conversation, ChatGPT can display cards with relevant information, such as weather forecasts, stock quotes, or sports results. Another feature is the ability to redirect complex queries to more powerful OpenAI language models. The voice model focuses on maintaining conversational naturalness, and if a question requires deep analysis or internet searching, the system activates the core language models, while the voice model continues the dialogue, and then voices the result.
OpenAI engineers also paid attention to the security of the new model. Additional moderation mechanisms have been implemented in the system, continuously monitoring both the incoming audio stream and the generated response. If potentially dangerous content is detected, the system can adjust the direction of the conversation or forcibly end the session. The senior version, GPT-Live-1, is available to paid ChatGPT users, while the simplified version, GPT-Live-1 mini, is intended for free accounts.
Information sources are a key element of any journalistic work, as they explain facts, lend credibility to the disseminated information, build trust, and contribute to a strict, transparent, and unbiased journalistic approach.
In the context of the controversy surrounding national exams, there is an indispensable source: the Ministry of Education. The Minister is obliged to promptly provide explanations for what happened, clarify the reasons for the failures, and present a clear action plan to minimize the consequences of the unsuccessful digitization process.
Important sources can include officials, experts, witnesses to events, or official documents. The value of each source is determined by their knowledge of the subject, authority in the field, and public reliability. When discussing exams, the testimony of teachers who participated in grading the papers is particularly important, as their accounts allow for the identification of shortcomings in a system that is far from ideal functioning.
An analysis of various news publications reveals numerous quotes pointing to this problem. Consequently, two main questions arise: who was responsible for developing and operating the platform used for digital assessment of papers, and what explanation does the Ministry of Education offer regarding such significant failures that affect thousands of students and their families?
In a Saturday publication by 'Público', the headline read: 'Exam digitalization cost more than seven million euros over three years.' In this article, journalist Andreia Sanchez used the Base Portal, which aggregates information on government contracts, to identify the organizations with which the Ministry of Education concluded agreements. She checked data up to 2023 and identified several companies, some of which had contracts for very large sums.
To find out who was involved in this process, the journalist interviewed the relevant managers. Some claimed they had no involvement in this work, while others referred to 'strict adherence to confidentiality obligations.' Although one can discuss the legitimacy of a private company's refusal to provide information of public interest, this freedom of judgment cannot be granted to the Ministry of Education, which, as the main body responsible for the education system, is obliged to provide all explanations. Comments are not enough; an official statement from the minister on this matter is required. The longer time passes, the more heated the controversy becomes, negatively affecting everyone. At stake is not only the future of many students but also trust in the university admission system.
