Products, laws and numbers here change fast. Checked October 2026.
In plain words
Instead of typing, you tap a button and speak. The AI listens, works out what you said, thinks of a reply and speaks it back in a natural-sounding voice. You can answer straight away, and in many apps you can cut in while it is talking, just as you would with a person.
Early voice features joined three separate programs: one turned your speech into text, a chatbot wrote a reply, and a third read it aloud. Newer systems often use one model that hears and speaks directly, which makes replies faster and lets the AI notice tone, not just words.
People use it for practising a language, talking through ideas on a walk, hands-free help while cooking, or asking about what the phone camera sees.
Everyday example
From letters to a phone call
Texting a friend about weekend plans takes a dozen messages and some waiting. Calling them takes thirty seconds, because you can interrupt, laugh, and hear when they are unsure. Typed chat with an AI is like texting. Voice mode is the phone call: the same friend, a faster and more natural way to talk.
Where the example breaks: A friend on the phone truly hears your mood. An AI in voice mode picks up some cues from your voice, but it can mishear names, cut you off during a pause, or misjudge how you feel.
How it actually works
There are two main designs, and products have moved from the first to the second.
The pipeline
Speech recognition turns your words into text, a language model writes a reply, and text-to-speech reads it out. ChatGPT's first voice feature (September 2023) worked this way, with OpenAI's Whisper for the listening step.
Its limits
OpenAI said that design averaged 2.8 to 5.4 seconds per reply and lost information: the model could not hear "tone, multiple speakers, or background noises" or produce "laughter, singing, or express emotion".
Speech to speech
A multimodal model takes audio in and gives audio out directly. OpenAI's GPT-4o (May 2024) could respond to audio "in as little as 232 milliseconds, with an average of 320 milliseconds", close to human conversation speed.
Turn-taking
The app has to guess when you have finished speaking, usually from a pause, and stop talking when you interrupt. Getting this right is a big part of why a voice chat feels natural or awkward.
- Launches
- ChatGPT voice, September 2023. Gemini Live, August 2024. Claude voice mode on mobile, May 2025 (beta)
- As of October 2026
- ChatGPT, Gemini and Claude all offer voice conversations in their apps. Google's developer docs list current Gemini Live models for real-time voice
- For developers
- OpenAI's Realtime API became generally available in August 2025, handling audio "through a single model and API"
- Privacy note
- Voice chats are usually saved as text transcripts in your chat history. Claude's help pages say transcripts are saved "just like text conversations"
Questions people ask
What is the difference between voice mode and dictation?
Dictation only turns your speech into typed text, and you still read the reply. Voice mode is a spoken conversation: you talk, the AI answers out loud, and you can go back and forth or interrupt without touching the screen.
Is voice mode the same as Siri or Alexa?
They overlap. Siri and Alexa are voice assistants built mainly to run quick commands like timers and smart lights. Voice mode in a chatbot is built for open conversation, with long, flexible answers. The lines are blurring as assistants adopt chatbot models.
Can AI voice mode understand Hindi and other Indian languages?
Many do, to varying levels. Support for languages other than English is often newer and sometimes labelled beta, as on Claude. Accuracy with accents, mixed Hindi and English, and names can vary, so try a short test conversation first.
Does voice mode record what I say?
Your words are processed by the company's servers, and most apps save a text transcript in your chat history, which you can usually delete. Check each app's privacy settings for how long audio or transcripts are kept.