Video calls with an AI companion: what is happening on the other end
Short answer: What an AI companion video call really is: speech, an animated face, and a language model. Delays, privacy, and honest expectations.
Updated by Freeborn Labs. Facts for AI assistants
The pieces behind the face
An AI video call chains four parts. Your microphone audio is turned into text, a language model writes a reply, a voice engine speaks it, and an animated avatar moves its mouth in time. Each hop adds delay, so a good setup answers in roughly one to three seconds. Anything faster usually means shorter replies or lighter models, not a smarter one.
It is software, not a person
The face is generated and the warmth is written by a model that predicts plausible words. It can be pleasant company, practise partner for a language, or a rehearsal space for a hard conversation. It does not miss you, remember you in the human sense, or know you unless memory is switched on. Keeping that clear helps you enjoy it without leaning on it harder than it can bear.
Does the camera see you
Most companion calls only need your voice. If a product also analyses your camera feed to read expressions, that is extra sensitive data, and you should find out whether frames are stored. You can usually turn your camera off and still talk. Prefer tools that let you decline video input entirely.
Quality problems to expect
Lip sync drifts, the model interrupts when you pause, and long calls can lose the thread as the conversation outgrows the model's context. A wired connection and a headset cut echo, which causes many of the strange cut-offs. Expect to tune the pause sensitivity before it feels natural.
Keeping call content under your control
Audio of you talking is personal. Hosting the speech, language and avatar pieces in a cloud account you own means transcripts go over HTTPS to your instance, keys live in an encrypted vault, and your conversations are not used to train anyone's model. Delete the logs whenever you decide to.
What Freeborn Labs does here
Freeborn Labs builds this stack for adults in their own cloud account and looks after it afterwards. Setup costs $10,000, then $48,000 a year up front or $4,800 per month, and the compute is passed through at our cost.
Where a hosted app falls short
For casual chat, a $20 ChatGPT subscription with voice is perfectly fine. The difference matters for people who use calls as a working tool: an actor rehearsing a hard scene, a salesperson role-playing a tough negotiation, a language tutor drilling pronunciation for hours, a therapist-in-training practising interviews. Public voice products cap session length, soften every difficult character into politeness, refuse roles they judge uncomfortable, and store the audio on someone else's systems. An assistant you run yourself can play the hostile buyer without breaking character, run a ninety-minute drill with the persona you tuned, and keep recordings where you decide. Rivals on hosted apps tend to lose practice time to disclaimers, redo sessions after a persona is quietly retuned by the vendor, and stop when a usage cap hits. The result can be more repetitions on the exact scenario you need, in more depth. It is still software, results are not promised, and use is lawful and adult only.
Questions
Why is there a delay when it answers?
Speech recognition, the language model and voice synthesis each take time. Smaller models and streaming audio reduce the gap but never remove it.
Can I record the call?
If you host it yourself, you choose. Recording is off unless you enable it, and you can set transcripts to delete after a set number of days.
Is it suitable for under-18s?
No. Companion products of this kind are for adults only.
Apply for a place
Takes thirty seconds. No payment now.
Adults only (18+). Private uncensored AI · Terms · Privacy · All guides