[DigitalToday reporter Lee Won-seon (이원선)] ChatGPT has applied its voice conversation model, GPT-Live, to make real-time dialogue sound more natural. Engadget reported on Aug. 9 that the model is built to let it speak while listening to a user.

ChatGPT's initial voice mode in 2023 worked by passing through three systems: speech-to-text, a language model and text-to-speech. The advanced voice mode introduced later moved to a single multimodal model, improving overall quality.

ChatGPT now uses GPT-Live for voice conversation processing. OpenAI described it as a "full-duplex" architecture. Earlier voice models sometimes replied before a sentence ended because they mistakenly judged a brief pause by the user as the end of an utterance.

GPT-Live is designed to reduce such problems. ChatGPT can intermittently offer short reactions such as "um" or "yeah" while the user is speaking, creating a flow closer to a real conversation.

Other models handle complex questions that require reasoning or research. GPT-Live can pass such requests in the background to models such as GPT-5.5 for processing. Generative AI has improved features, accuracy and response speed with each new model, but the limitation remains that answers should not be treated as an absolute source of information.

ChatGPT also has another new feature. It can change not only the content of an answer but also the appearance of the text. TechRadar reported that it can create decorative text using Unicode character styles without installing separate font files.

The feature does not work by installing font files like a word processor. ChatGPT rewrites sentences into Unicode character forms. Users can enter the desired phrase in a new chat window and specify a concrete style such as speech-bubble lettering, a gothic look or inverted letters. They are more likely to get the desired result by asking for a "specific Unicode style" than simply requesting a "different font."

It can also be used to change the tone of short phrases. Applying it to phrases such as birthday messages, party invitations, year-end greetings or social media posts can give a different impression with the same content. Text created by ChatGPT can also be copied and pasted into other apps.

There are limits as well. Unicode-based text displays properly only in environments that support it. It works without major issues in most modern apps, but on some websites it may appear as empty squares or be replaced by other symbols. Highly decorative styles can reduce readability and may also be inconvenient for accessibility tools such as screen readers.

For this reason, it is appropriate to limit such decorative text to titles or short emphasis phrases. ChatGPT noted that while it can convert an entire long piece of writing, its usefulness declines if the original exceeds a range that is easy to read. It is not a feature that greatly boosts productivity, but it can be used to make social media posts, congratulatory messages and invitations a little more noticeable.

The key change is that the voice interface has begun to handle conversation timing, going beyond a simple readout function. It also shows that such features allow generative AI to adjust not only sentence content but also the form of expression.