Realtime voice agentspeech in, speech out

A voice agent that talks like a person.

Real speech-to-text in, real speech out, barge-in when you interrupt, engineered around one number: perceived turn latency. Runs on the browser's own Web Speech API.

No sign-up, no API key, no server. Audio is never stored.

Sample callRestaurant booking from the console

Interrupt the agent while it speaks. If it is not speaking, the barge-in waits for its next reply.

  1. Incoming call, restaurant booking.
  2. Agent: Thanks for calling Bistro Verde — how can I help with your reservation?
  3. Caller: A table for two tonight
  4. Agent (tool call check_availability, party 2 · tonight, window 6–9 pm, slot 7:15 pm): I can do 7:15 tonight for that party size. Want me to lock it in?
  5. Caller: Any outdoor seating?
  6. Agent (tool call check_availability, party 2 · tonight, patio · 6–9 pm, slot 8:45 pm): I have a patio table at 8:45 pm tonight for that party size. Shall I, cut off. Caller barges in: Can I add a high chair?
  7. Turn latency after the barge-in: 398 milliseconds: endpointing 30, speech-to-text 90, reasoner 158, text-to-speech 120.
  8. Agent: Of course, I'll note a high chair on the booking. Anything else?
  9. Caller: Book it, please.
  10. Agent: Booked! The patio table at 8:45 pm tonight. You'll get a text confirmation shortly.
Line audioinked as it streamscalleragent1 second

What it does

One agent, four lines. Answers come from its skills.

Pick up any of four calls in the console. The agent listens, decides whether it needs a skill, calls it, and speaks the answer while you can still cut in. Every fact it states comes from a small table that lives in your browser.

Line 11 skill

Restaurant booking

Thanks for calling Bistro Verde — how can I help with your reservation?

  • check_availability(party, date, window, area?)Finds open tables that seat the party inside a time window, indoors or on the patio, nearest to the time asked for.reads 10 table slots

CallerA table for two tonight

check_availability → slot 7:15 pm

AgentI can do 7:15 tonight for that party size. Want me to lock it in?

Line 24 skills

Support agent

Hi, thanks for calling. How can I help with your order today?

  • order_status(order)Reads where an order is and when it should arrive.reads 1 order on file
  • cancel_order(order)Cancels an order only if it has not shipped yet.reads order status
  • start_return(order)Checks the 30-day return window and opens a return.reads order date
  • store_hours()Reads the support desk's opening hours.reads hours table

CallerWhat's my order status?

order_status → shipped · eta Thu

AgentYour order shipped this morning and should arrive Thursday. Want me to text you the tracking link?

Line 32 skills

Coffee barista

Welcome in! What can I get started for you?

  • menu_lookup(item)Confirms a drink is on the menu and reads its prices.reads 6 drinks, 3 sizes
  • ring_up(items)Totals the order from the menu and milk prices.reads menu prices

CallerA large oat latte please

menu_lookup → in stock · from $3.90

AgentGreat choice — one large oat latte. Anything else, or shall I ring you up?

Line 4no skills

Interview coach

I'll play the interviewer. Say “I'm ready” and I'll ask you a question.

  • Pure conversation, no tools. With nothing to look up, the turn is the bare pipeline: listen, think, speak.

CallerI'm ready

AgentHere's one: describe a time you disagreed with a teammate. Use situation, action, and result.

How it works

Listen, think, speak, with no dead air between.

One turn of the call in the strip above, written out cue by cue. The timecodes and milliseconds are the simulation's own numbers for this sample call.

Converse · sample call · restaurant bookingTwo turns and a barge-in
00:05.3CallerA table for two tonight.Words arrive as speech-to-text partials, outlined in pen until the final result inks them solid.
00:06.6EndpointerSilence. The caller has stopped; the clock starts.30 ms
00:06.7STTfinal: “a table for two tonight”90 ms
00:06.7Reasonercheck_availability(party: 2, date: tonight, window: 18:00–21:00)Reads the line, picks the skill, fills its arguments. The skill finds 2 open tables and holds the nearest.195 ms
00:07.1AgentI can do 7:15 tonight for that party size. Want me to lock it in?The first word is on the line while the rest is still streaming.120 ms
From the caller's last word to the agent's first: the one number this agent is built around.435 msturn latency

Turn 2 and the barge-in

00:11.7CallerAny outdoor seating?
00:12.7Reasonercheck_availability(party: 2, date: tonight, window: 18:00–21:00, area: patio)The same skill, filtered to the patio. The 8:45 pm patio table it finds replaces the 7:15 pm hold.189 ms
00:17.0Caller
(over the agent)
Can I add a high chair?
00:17.1AgentI have a patio table at 8:45 pm tonight for that party size. Shall I— hold it for you?Playback is cancelled on the word it reached. The rest is dropped, never queued.
00:18.9AgentOf course, I'll note a high chair on the booking. Anything else?398 ms after the caller stopped

Architecture

Four stages in a line, and a wire back for interruptions.

Each stage streams into the next instead of waiting for it to finish. When the caller talks over a reply, the loop runs backwards: playback stops on the spot and the turn starts again from listening.

Caller speaks
30 ms

Endpointing

Decides the caller has finished, from the silence after their last word.

Web Speech end of speech
90 ms

Speech-to-text

Partials stream while the caller talks; the final result closes the line.

SpeechRecognition, interim results
140+ ms

Reasoner

Reads the line, calls a skill when it needs a fact, and starts the reply.

check_availabilityorder_statuscancel_orderstart_returnstore_hoursmenu_lookupring_uprule-based agent, typed skills
120 ms

Text-to-speech

Speaks the reply as it streams, and stops the instant the caller cuts in.

speechSynthesis, cancel() on barge-in
Caller hears
barge-in: caller talks → playback cancelled → listening again
The budget, sample turn
435ms
The turn from the strip above: caller's last word to the agent's first, with one skill call.
ENDPOINT 30STT 90REASONER 195TTS 1200100200300435caller stopsfirst word
Measured or modelled
Endpoint and STTmeasured when you speakThe console times the gap between Web Speech's end of speech and its final result. The sample call uses 30 + 90 ms.
ReasonermodelledA small rule-based agent answers in microseconds, so it is timed like a streaming model: 140 ms plus 3 ms per word heard and 40 ms per skill call.
Text-to-speechmeasured with voice onFrom speechSynthesis.speak() to the utterance's start event. Silent turns use 120 ms.
Audionever storedSpeech goes to your browser's own recognition service (in Chrome, Google's) and is dropped when the turn ends. This site has no server to send it to.

The stack

Everything runs in your browser.

No server sits between you and the agent: the page is the whole product, from the microphone to the speaker.

Converse No. 1Credits
Ears
Web Speech API: SpeechRecognition with interim results
Voice
speechSynthesis, cancelled the moment you barge in
Script
Rule-based domain agents with typed skills over in-browser tables
Timing
A stage latency model, checked against performance.now()
Pencils
Next.js 16, React 19, TypeScript
Inks
Hand-ruled SVG drawn by code, seeded so the server and the browser draw the same frame
Letters
Fraunces, Libre Franklin, IBM Plex Mono
Colours
None
Proofs
Vitest, ESLint, Lighthouse
Printing
Netlify. No backend, no API key, nothing stored
Created by
Riken Patel
The Converse owl in a headset, ready to take the call
Agenton the line

Your turn. Pick up the line and talk over it.

Launch the console

Bring a microphone, or type. No sign-up, no API key.