How does voice AI work?
A voice AI system has three jobs: work out what was said, decide what to say back and say it. A chained, or cascaded, system gives each job its own component. Automatic speech recognition (ASR) turns the audio into text, a language model writes a reply and text to speech (TTS) reads it out. A speech-to-speech model works on the audio itself, so one model hears the question and speaks the answer.
OpenAI’s guide to voice agents sets out the trade-off. A chained pipeline lets you inspect or change the text between stages and replace each part separately. A speech-to-speech session uses one model to interpret audio, decide what to do and respond in speech.
Example: a visitor asks aloud, “Where do I change my billing email?” The system recognises the words, finds the answer in the product’s help articles and speaks the steps back in a chosen voice.
Web browsers have speech features of their own. The Web Speech API gives a page speech recognition and speech synthesis, though many voice agents use a cloud voice model instead.
Why does voice AI matter for a SaaS website?
Some questions are easier to say than to type, such as describing a setup problem step by step or asking which plan suits a team. Voice AI lets a visitor ask on the page they’re reading and hear a spoken answer, without searching help articles or waiting for an email.
The hard parts are the same in every product: answers have to come from accurate information, the system has to take turns naturally, and people should know they’re talking to software. In the EU, Article 50 of the AI Act requires AI systems that interact directly with people to be designed so those people are told they’re dealing with an AI, unless that’s obvious.
How does Kawl Now use voice AI?
Kawl Now is a browser voice AI agent for SaaS websites. A visitor clicks the widget, types what they need, allows the microphone and talks, and a support or sales agent answers from knowledge the SaaS team has approved.
The voice runs on OpenAI’s GPT-Live model over WebRTC. Audio goes between the visitor’s browser and OpenAI, and Kawl Now saves a text transcript, not a recording. Agents greet callers in English and speak in one of two voices, Marin or Cedar.
Kawl Now uses voice AI for conversations on your website. It has no phone number for customers to call, doesn’t answer phone calls and isn’t a text-to-speech or voice cloning tool.
Common questions
Is voice AI the same as text to speech?
No. Text to speech is one part of voice AI: it turns written text into spoken audio. A voice AI system also has to understand what a person says and decide how to reply, using speech recognition and a language model, or a speech-to-speech model that does both.
What is the difference between voice AI and a voice assistant?
Voice AI is the technology. A voice assistant is one product built with it, usually a general helper on a phone or smart speaker that answers questions and carries out commands for its owner. A website voice agent is another, built to answer one company’s customers about its products.