To add a voice assistant to your website, you need three things: a way to hear the visitor (the microphone and speech recognition), something that decides the answer (usually an AI model reading your own content), and a voice to speak it back. You can build those with browser APIs, use a voice-agent platform, or turn on voice in a chat widget.
This guide compares the three paths, lists the technical requirements that trip people up, and shows what each one costs.
Key Takeaways
- A website voice assistant is a microphone button, a speech-to-text step, an answer engine and a text-to-speech voice. Every option on the market is some mix of those four parts.
- Microphone access only works on HTTPS pages, and the visitor must allow it once.
- Browser speech recognition is free to call but not supported everywhere, and in Chrome the audio goes to a server anyway.
- Always keep a text fallback. Plenty of visitors can't or won't speak out loud.
- The part that decides quality is the answer, not the voice. A natural voice reading a wrong answer is still a wrong answer.
What a website voice assistant is
A website voice assistant lets visitors ask a question out loud on your site and hear the answer spoken back, without calling a phone number or installing anything. It sits in the page, usually as a button inside a chat widget.
It is different from three things it often gets confused with:
| Website voice assistant | Phone AI receptionist | Text-to-speech reader | |
|---|---|---|---|
| Where it runs | In the browser, on your page | On a phone number | In the browser |
| What the visitor does | Taps and talks | Calls | Listens |
| Understands questions | Yes | Yes | No, it only reads text aloud |
| Typical use | Product questions, help while browsing | Bookings, calls when you're busy | Accessibility, reading articles |
If what you actually want is to answer calls, see our breakdown of AI receptionist costs instead.
Three ways to add voice to your website
There are three realistic paths, and the right one depends on how much you want to build.
| Build it with browser APIs | Voice-agent platform | Chat widget with voice | |
|---|---|---|---|
| Examples | Web Speech API plus your own backend | Retell, Vapi, ElevenLabs Agents | Conecto, and other widgets that add a voice mode |
| Setup | Weeks of engineering | Days: design the agent, embed its web SDK | Minutes: paste a snippet, switch voice on |
| Answers from your content | You build retrieval yourself | You connect a knowledge base | Uses the widget's existing knowledge base |
| Text fallback and inbox | You build them | Usually separate | Included, same conversation |
| Cost model | Your AI and hosting costs | Per minute | Plan price plus usage |
| Best for | Teams with engineers and unusual needs | Voice-first products, custom flows | Businesses that want voice on top of support chat |
Path 1: Build it with browser APIs
Browsers include the Web Speech API: SpeechRecognition turns speech into text, and speechSynthesis reads text aloud. You connect them to an answer engine of your own, typically an AI model with your help content as context.
What to know before you start:
- Support is limited. MDN marks
SpeechRecognitionas "Limited availability" because it does not work in some of the most widely used browsers. MDN, "SpeechRecognition" - It is not offline by default. In Chrome, speech recognition on a web page uses a server-based engine, so the visitor's audio is sent to a web service for processing. Newer browsers let you request on-device recognition with
processLocally. MDN, "Using the Web Speech API" - The built-in voices sound built in.
speechSynthesisuses the device's voices, which vary by operating system and rarely sound like your brand.
This path costs the least in fees and the most in engineering time. It fits a team that wants full control and has someone to maintain it.
Path 2: Use a voice-agent platform
Voice-agent platforms handle the listening, the model and the voice, and give you a web SDK to embed. You design the agent's instructions and connect your knowledge.
They price per minute of conversation. Retell lists $0.07 to $0.31 per minute for voice agents, depending on the model and voice you pick. Retell AI pricing Vapi breaks the cost into parts you choose separately: transcription, the language model, the voice and telephony, each billed per minute. Vapi pricing
This path fits products where voice is the main experience. The trade-off is that the voice agent usually lives apart from your chat, your inbox and your customer records, so you connect those yourself.
Path 3: Turn on voice in your chat widget
Some chat widgets now include a voice mode: the same assistant that answers typed questions can listen and speak. Setup is the widget's snippet plus a switch, and spoken conversations land in the same inbox as typed ones.
In Conecto, voice is part of the AI voice widget: visitors tap the microphone in the chat, the AI answers out loud from the same knowledge base as text chat, and your team sees the transcript in the shared inbox. Spoken replies use the same AI credits as text replies; the arithmetic is on the pricing page.
This path fits businesses that already need website chat and want voice for the visitors who prefer it, without a second system.
Technical requirements for any website voice assistant
Whichever path you choose, these requirements apply.
- HTTPS. Browsers only allow microphone access on secure pages;
getUserMedia(), the method that asks for the microphone, is available only in secure contexts. MDN, "MediaDevices: getUserMedia()" - A clear permission moment. The browser shows its own "allow microphone" prompt. Ask only after the visitor taps the voice button, never on page load, or most people will block it.
- A text fallback. Some visitors are in open offices, on trains, or can't speak. The same assistant must work by typing.
- Answers from your content. The assistant should answer from your help center, product pages and policies, and say "I don't know" when those don't cover the question.
- A way to a person. Voice conversations need the same human handoff as chat, with the transcript attached.
- Transcripts and privacy. Save what was said so your team can review it, tell visitors that the conversation is recorded as text, and cover it in your privacy policy.
How to add a voice assistant to your website, step by step
- Pick your path from the comparison above. If you already run a chat widget, check whether it has a voice mode before building anything.
- Prepare the answers. Make sure your help content covers the top 20 questions visitors ask. A voice assistant exposes gaps faster than text, because a wrong spoken answer is harder to skim past.
- Install it on one page first. Start with the page that gets the most questions, often pricing or a product page.
- Write the first spoken line. Keep it short and specific: "Hi, ask me anything about delivery or sizes." Long spoken introductions make visitors wait.
- Test on real devices. Try iPhone and Android, Chrome and Safari, with and without headphones, in a quiet and a noisy room.
- Read the first 50 transcripts. Fix the answers that went wrong, then roll out to the rest of the site.
What makes a voice assistant feel good to use
- Fast first words. Visitors wait a few seconds at most for a spoken reply. Streaming the answer sentence by sentence beats waiting for the whole paragraph.
- Short answers. A spoken answer should be two or three sentences. Put the detail in a link or on the screen.
- Interruptible. Visitors should be able to talk over the assistant to correct it.
- Shows as well as tells. A voice answer about "the blue button at the bottom" is weaker than an assistant that highlights the button. That is the idea behind Beside, which points at the part of the page it is talking about.
Common mistakes
- Voice without a knowledge base. A general model answering from memory will invent delivery times and prices.
- Asking for the microphone on page load. Most visitors decline and never see the prompt again.
- No text option. Voice-only locks out everyone who can't talk right now.
- Long spoken paragraphs. Listening is slower than reading. Cut answers in half.
- No transcripts. You can't improve what you can't read.
Frequently asked questions
Can I add a voice assistant to my website for free?
You can prototype one with the browser's Web Speech API at no fee, but it won't work in every browser, and you still need an answer engine behind it. Production voice assistants cost either engineering time or a per-minute or plan fee.
Do visitors need to install anything to talk to a website voice assistant?
No. It runs in the browser. The visitor allows microphone access once when they tap the voice button.
Does a website voice assistant work on mobile?
Yes, on HTTPS pages in modern mobile browsers, as long as the visitor allows the microphone. Test on both iOS and Android, since microphone and audio behavior differ.
How much does a voice assistant for a website cost?
Platforms charge per minute, for example $0.07 to $0.31 per minute at Retell as of September 2026. Chat widgets with voice include it in a plan and meter the AI usage. Building it yourself shifts the cost to development and AI model fees.
Is a voice assistant good for accessibility?
It helps visitors who find typing hard, but it can't be the only way in. Keep full keyboard and text access, and don't rely on sound alone to convey information.
Where to start
If you already need live chat and AI answers, start with a widget that includes voice and turn it on for one page. If voice is your product, compare the per-minute platforms. You can try the voice path in Conecto on any paid plan: see how the AI voice widget works or start a workspace.