How does speech-to-speech AI work?
A conversational speech-to-speech model works on audio directly. It hears the speaker, decides what to say and generates the spoken reply itself, instead of passing text between separate systems. OpenAI describes its Realtime API this way: one model interprets audio, decides what to do and responds in speech.
The alternative is a chained, or cascaded, pipeline: speech to text, then a text model, then text to speech. Each stage can be inspected or replaced. OpenAI’s example is an app that stores the transcript, runs policy checks before the text agent responds and only generates speech once it has an approved answer.
Two other uses share the name. Speech-to-speech translation turns speech in one language into speech in another, which OpenAI offers as a dedicated live translation session. Voice conversion keeps the words and delivery but changes the voice: ElevenLabs’ API reference, read on 8 October 2026, serves its voice changer from a speech-to-speech endpoint.
Example: asked “Can I pause my plan?”, a speech-to-speech agent produces the spoken reply directly, while a chained agent first writes the answer as text and then has it read out.
Why does speech-to-speech matter for a voice agent?
For live conversation, fewer hand-offs between separate systems mean fewer places for delay to build up. A chained pipeline gives more control: you can check the text before it’s spoken, keep an exact record of it and choose each component separately.
Neither choice decides what the agent says. That comes from its instructions and the information it can search, so test its answers the same way whichever architecture it uses.
How does Kawl Now use speech-to-speech?
Kawl Now runs on OpenAI’s GPT-Live, a voice model that hears the caller and speaks its replies itself, so there’s no separate speech-to-text or text-to-speech stage for you to configure. OpenAI describes GPT-Live as full duplex: it can listen while it speaks.
Answers to product questions still come from text, because each agent is told to search its channel’s approved knowledge. The voice service also returns a running transcript of both speakers, which the widget shows as live captions and Kawl Now saves.
Kawl Now doesn’t offer speech translation or voice conversion. Agents greet callers in English, with no language setting, and speak in one of two preset voices.
Common questions
Is speech-to-speech better than speech to text plus text to speech?
Neither is better for every job. Speech-to-speech suits live conversation because one model handles the whole exchange. A chained pipeline suits teams that need to inspect, store or approve the text before it’s spoken, or to choose each component themselves.
What does S2S mean in AI?
S2S is short for speech to speech: AI that turns spoken input into spoken output, usually with one model working on the audio itself, whether to hold a conversation, translate between languages or change the voice of a recording.