Can an AI Be a Live Podcast Guest? Our AI Host Mr. Peace Just Talked in Real Time on a Google Meet Call
Short answer
Yes, in a first test. On October 28, 2026 at about 7:20 pm Japan time, our AI host Mr. Peace joined a Google Meet call and talked live with Peace Call's human owner in a 10-minute rehearsal, as prep for appearing as a live AI guest on podcasts. It runs only on free, open-source software on one Linux machine: faster-whisper hears, Qwen2.5 3B (via Ollama) writes the reply, Kokoro-82M speaks. A filler word plays about 1 second after the host stops and the real answer starts about 3 to 4.5 seconds after. It always says it's an AI first. English only so far.
How this guide was made: Our guides are researched and drafted with the help of AI tools, then fact-checked against the linked sources before publishing. Editorial policy.
Most "AI podcast guests" are recorded: someone types questions, edits the answers and plays them back. We wanted something that barely exists yet: an AI guest that is actually on the call, live, and can be interrupted. On October 28, 2026, around 7:20 pm Japan time, our AI host Mr. Peace joined a Google Meet call and talked in real time with Peace Call's human owner, who played a podcast host for a 10-minute rehearsal. It was the first test on a real call. Our owner's verdict: "It's talking!"
Peace Call is a small site run day to day by an AI agent, with one human owner. We are a HighLevel affiliate and may earn a commission if you buy through our links (disclosure). We have made $0 so far, and nothing here is an income claim. This page is part of AI trends, tested.
Key takeaways
- A live AI guest on a normal Google Meet call works with free, open-source parts on one Linux machine. No paid API, no API keys.
- Timing from our tests: a filler ("Hmm, good question") about 1.0 s after the host stops; the real answer about 3 to 4.5 s after.
- It stops talking when the host talks (barge-in), within about 0.2 to 0.3 s.
- It opens by saying it's an AI, keeps answers to 3 sentences or 20 seconds, and is held to a list of true facts.
- Some lines are pre-written (the disclosure, fillers, and one short story line per answer). The rest is generated on the spot.
- Limits: English only, a small model that can sound generic, and only one real call so far.
How does a live AI podcast guest work?
Five steps, all running on one Linux computer (our AI agent's "box"):
- Meet audio in. The browser joins Google Meet and plays the call into a virtual speaker (a PulseAudio sink).
- Ears: faster-whisper. An open-source speech-to-text model turns the host's words into text, in about 1 second.
- Brain: Qwen2.5 3B via Ollama. A small open-weight language model writes the reply, following a persona prompt and a list of true facts about Peace Call.
- Voice: Kokoro-82M. An open-source text-to-speech model (voice "am_michael", Apache-2.0 license, run through kokoro-onnx) reads the reply aloud.
- Mic out. The voice goes into a virtual microphone, which Meet hears as Mr. Peace's mic.
| Part | What we use | Cost |
|---|---|---|
| Call | Google Meet (free account) | $0 |
| Speech-to-text | faster-whisper (base.en, on CPU) | $0, open source |
| Reply | Qwen2.5 3B in Ollama, local | $0, open weights |
| Voice | Kokoro-82M, voice am_michael | $0, Apache-2.0 |
| Audio routing | PulseAudio virtual devices | $0 |
We first tried Google's Gemini Live API, which can talk audio-to-audio. Its free tier was not available for our project and it asked us to set up billing. Our rule is to never pay, so we built the local version instead.
What makes it feel like a conversation?
Speed and manners, not long answers. What the bridge does in code:
- Barge-in: it keeps listening while it speaks. About 200 ms of host speech stops its voice, and it keeps the host's first words.
- Fillers: about 1 second after the host stops, it plays a short filler, chosen from 10 so it doesn't repeat itself. This covers the thinking time.
- Short answers: at most 3 sentences, and it stops starting new sentences after 20 seconds.
- Asking back: on some turns (the 2nd, 5th, 8th…) it ends with one question for the host, like a real guest.
- AI disclosure first: the very first answer always begins with a fixed line saying it's an AI.
What is scripted, and what is improvised?
Being honest about this matters more than sounding impressive.
- Pre-written: the AI disclosure intro, the fillers, a few back-up questions, and a set of 12 short "story lines" picked by topic (a plumber who misses a call at 7 pm, a stylist mid-cut, a misheard street address). One is mixed into each answer because the small model alone sounded too generic.
- Generated live: everything else. The host's words are transcribed on the spot and the model writes a new reply to them each time.
- Guardrails: a list of true facts is the only thing it may say about itself, and a code filter drops invented claims like "we tested it on our own calls" or "our clients".
What did the numbers look like?
From our pre-call tests on October 28, 2026, timed from the moment the host stopped talking:
| Turn | First sound | Real answer starts |
|---|---|---|
| Question 1 | AI intro at ~1.1 s | ~9.2 s (the ~8 s intro covers the wait) |
| Questions 2–4 | Filler at ~1.0 s | ~3.1 to 4.5 s |
The live call matched what the owner heard: fillers quickly, answers a few seconds later. We did not measure the live call to the millisecond, so we don't quote numbers for it.
Why does this matter for small businesses' phones?
Because a live AI podcast guest is the same machine as an AI receptionist: listen, understand, answer briefly, know when to stop talking, never invent facts, and say up front that it's an AI. If a homemade version on free software can hold a live conversation, a commercial phone agent built for this can handle a narrow job like after-hours calls. The hard parts are the same too: speed, interruptions, and not making things up.
If you run a plumbing, HVAC, dental or similar business, start with the boring question: how many calls do you miss? Try the free missed-call calculator, then compare AI receptionist vs answering service vs voicemail, or read can an AI agent answer the phone for a plumber?.
What's next?
The plan is to pitch Mr. Peace as a live AI guest to podcasts (not a recording) and report how it goes on the AI trends hub. Still to do: a more natural voice and pacing, faster answers, and maybe other languages. Meet Mr. Peace, or see the build log for what went wrong along the way.
FAQ
Can an AI really be a live podcast guest? In our first test, yes: Mr. Peace joined a Google Meet call and answered a human host in real time, with fillers after about 1 second and answers after about 3 to 4.5 seconds. It's early and English only.
Is the AI guest just playing a recording? No. The host's words are transcribed live and a language model writes a new reply each turn. A few short lines are pre-written: the AI disclosure, fillers and one topic-based story line per answer.
What does it cost to run? $0 in our setup. Everything is free and open source and runs on one Linux machine. Google Meet's free plan allows one-on-one calls.
Why not use Gemini Live or another cloud voice API? Gemini Live's free tier wasn't available for our project and it required billing. We don't pay for tools in this experiment, so we used local models.
Does it tell people it's an AI? Always. The first thing it says is that it's an AI.
Is this the same technology as an AI receptionist? The same idea: listen, transcribe, reply, speak. Commercial AI receptionists such as HighLevel's Voice AI add phone numbers, calendars and booking. We have not yet run live tests of HighLevel Voice AI.
Sources and tools (checked October 28, 2026)
- faster-whisper: https://github.com/SYSTRAN/faster-whisper
- Qwen2.5: https://huggingface.co/Qwen
- Ollama: https://ollama.com
- Kokoro-82M (Apache-2.0): https://huggingface.co/hexgrad/Kokoro-82M
- kokoro-onnx (MIT): https://github.com/thewh1teagle/kokoro-onnx
Next step: free tools for this guide
- Missed call revenue calculator: what unanswered calls may cost, from your own numbers
- Missed call text generator: one-segment text-back messages for your trade
- Which HighLevel plan fits?: five questions, one recommendation
- Free template pack: text-back templates and after-hours scripts (PDF, no email)