Skip to content
Seller Lab

Seller Lab 1 · A hands-on lab for new LiveKit sellers

Why real-time voice AI is hard

Juniper is a voice agent that answers the phone for a made-up restaurant. Each experiment plays it the same recorded caller twice and changes one setting. You predict what will happen, hear the difference, try it with your own voice, and leave with the question to ask a customer.

Start the labHow it’s built

An open-source project by Tony Morinello

Experiment 1 · Turn detection

Same caller, two settings

  • Caller
  • Juniper speaking
  • Thinking
  • Cut off
  • Both talking

Baseline · silence ends the turn

Before its first answer, Juniper heard “I'd like a table for”. It started answering 3 times and was cut off twice.

Enhanced · LiveKit’s turn detector

Before its first answer, Juniper heard “I'd like a table for four people on Friday at seven.” It started answering once and was never cut off.

Read the Baseline call
  1. Caller: I'd like a table for
  2. Juniper: How many people will be in your party?cut off
  3. Caller: four people on
  4. Juniper: Which day would you like to join us?cut off
  5. Caller: Friday at seven.
  6. Juniper: And may I have a name for the booking?

Recorded on 14 Nov 2026: a real call in a LiveKit Cloud room, with the lab’s agent and models. You hear what the caller heard, the test clip and Juniper’s voice.

Open Experiment 1 in the lab

What’s inside

Two things to try, and a pilot that checks whether they teach

The lab

Make it your own, then four experiments

Rename the restaurant and give Juniper a house rule. Then turn detection, interruptions, background noise and model speed: each one runs the same test caller on both settings before you try it with your own voice.

Headphones · Chrome or Edge on a laptop

Open the lab

The practice call

A discovery call with Priya, an AI prospect

Eight to ten minutes with the CTO of a 12-person startup. A judge scores the call on six items, and every point needs a quote from the call. Make up a technical answer and the call fails. Then a coach talks you through it.

Pass at 8 of 12 · No made-up answers

Call Priya

The pilot

A check before and after, with consent first

Five questions and three ratings before the lab, and the same again after the practice call. Nothing is recorded until a participant agrees. Each session gets its own page of results, which can be deleted.

Four steps · About 50 minutes

How the pilot works

This is Lab 1 of a program

Getting sellers ready takes a curriculum, certification and measurement, not one lab. So there’s a design for the rest: what to find out first, five more labs, and what the pilot already measures.

See the program

What the live runs showed

Same clips, two runs per setting, on LiveKit Cloud

ExperimentBaselineEnhanced
1. Turn detectionBoth runs jumped in after “I’d like a table for” and answered at each pause.Both runs heard the whole sentence as one turn and answered once.
2. InterruptionsBoth runs stopped at each “uh-huh” and picked the sentence back up about two seconds later.Both runs read on through both “uh-huh”s and stopped only for the vegetarian question.
3. The coffee shopEvery run heard the next table as part of the call: 11 or 12 words nobody said, like “No way. Did you see the game last night?” And Juniper answered them.With voice isolation, every run heard only the caller, and Juniper went straight on to ask for a name.
4. The pausefast vs reasoning model0.54 to 0.69 seconds before the answer; the LLM’s first words took 0.12 to 0.17 of it.1.77 to 2.23 seconds; the LLM’s first words took 1.32 to 1.82 of it.

A cheaper judge

On two saved practice calls, scored twice each, gpt-5.4-mini reached the same verdicts as gpt-5.5. It takes 2 to 4 seconds instead of 4.5 to 6.5, and costs about a tenth as much. So it’s the judge now.

The hosted lab, end to end

The cloud agent joined within 8 seconds. A typed practice call came back scored, with a quote and a time for each point, within 15 seconds of hanging up.

How it’s built

One Python worker plays every voice in the lab

  1. BrowserNext.js pages and livekit-client
  2. Cloudflare WorkerAccess code, daily cap, a 2-minute join token
  3. LiveKit CloudThe room, and the agent dispatched to it
  4. AgentPython on LiveKit Agents 1.8: Juniper, Priya, the coach
  5. ModelsDeepgram Nova-3 and Aura-2, LLMs on LiveKit Inference

Settings switch on a live call

Turn detection changes on the running session with update_options. Interruption settings are fixed when an agent is created, so that switch hands off to a fresh Juniper that keeps the conversation.

Both settings hear the same audio

Test-caller clips play into one published microphone track through a Web Audio mixer. Your own microphone joins the mix only when you try it yourself.

The judge has to show its work

A point needs a quote that’s really in the call, matched word for word. A made-up technical answer zeroes technical honesty and fails the call, whatever the total. Ten tests cover these rules.

A public demo that stays cheap

Rooms need the access code, checked in constant time. At most 40 rooms a day; join tokens last two minutes; calls end on their own.

Read how it’s built