Voice Mode

Voice Mode refers to the voice feature of an AI assistant such as ChatGPT. Instead of typing a question, you simply say it out loud. The program listens, thinks, and responds with an artificially generated voice. The conversation runs much like a phone call: you take turns speaking without having to press a talk button. The term originates from OpenAI, but is now used generally for such voice modes. It always refers to the same idea: the same AI, just with ears and a mouth instead of a keyboard.

Why speaking lowers the barrier

Typing is slow. Most people speak about three times as fast as they can write on a phone. Anyone who is out and about, cooking, or driving doesn’t have a free hand anyway. Voice Mode makes the AI usable precisely in these situations.

There is also a group of users for whom text is a real barrier. Children, older people, and those with visual impairments or reading difficulties often get along noticeably better with speech. Learning a foreign language is also easier when you can actually talk and hear a response.

Economically, voice mode is interesting for providers because it puts their assistants in direct competition with Siri, Alexa, and Google Assistant. These older voice assistants only understand a limited number of commands. A voice mode based on a large language model, on the other hand, can talk freely about almost any topic. That is precisely why voice is currently considered one of the most important battlegrounds among the major tech companies.

From the detour through text to direct listening

The older design consists of three separate programs in a chain. First, a speech recognizer converts what was said into text. This text is then fed to the actual language model, i.e. the AI that formulates the answer. Finally, a third program reads the answer aloud using a synthetic voice.

This chain works, but it has two weaknesses. It costs time, because each station waits until the previous one is finished. And it discards information: the text only records what was said, not how. Whether someone is whispering, laughing, or sounds annoyed is lost.

Newer voice modes therefore work end-to-end. The model takes in sound directly and generates sound directly as a response, without the detour through written text. This reduces the delay to a fraction of a second, similar to a real conversation. In addition, the model can perceive tone of voice and produce it itself, for example speaking faster or answering in a whisper. It can also be interrupted mid-sentence, because it keeps listening while it speaks.

Voice modes in phones, cars, and debate

You most commonly encounter Voice Mode in the apps of ChatGPT, Google Gemini, and Microsoft Copilot. There it is usually found as a microphone or waveform icon. Apple has rebuilt its Siri with similar technology, and car manufacturers are building such assistants into their onboard computers. Customer service hotlines are also increasingly using the technology to answer calls without humans.

In the news, the term often comes up in connection with controversy. A well-known case was OpenAI’s “Sky” voice in 2024, which reminded many people of the actress Scarlett Johansson and was subsequently removed. A second recurring topic is data privacy, since voice recordings are more personal than text and are sometimes processed on foreign servers.

A common misconception is that a voice mode makes the AI smarter. That’s not true: behind it is the same model as in text chat, so it can just as easily make things up. But because a fluent, friendly voice sounds more convincing than a block of text, it becomes harder to stay skeptical. Important information should therefore be checked even when it sounds very confident.