Compagnie is an AI conversational companion designed for French speakers in Québec, France, Belgium and Switzerland. Unlike our other case studies, this is not a client project: Compagnie is a product that Voxivra designed, built and operates every day. It runs, at consumer scale, the same building blocks we deploy for our clients: telephony, real-time voice, language models and integrations.
Context
Most conversational assistants live in a chat window and forget everything between sessions. Compagnie takes a different approach: an ongoing relationship with a companion that keeps its personality and voice, remembers what matters to the person, and stays reachable through the most natural channel, including the phone.
The service targets a broad audience, including people who are not comfortable with apps. A simple phone call or text message had to be enough to start a conversation.
Challenge
- Natural voice conversation: short response times, smooth turn-taking, a stable and recognizable voice.
- Many channels, one relationship: app, in-app voice call, calls from the user’s own phone and SMS, without switching channels “resetting” the companion.
- Useful, controllable memory: remember tastes, projects and important people, separate real life from a shared fictional universe, and let users correct or delete each memory.
- French that sounds right: understand how users actually speak, in Québec as in Europe, rather than translating an English-first product.
- Privacy: encrypted exchanges, storage in Canada, no advertising use of conversations.
Solution
An integrated telephony layer
Users can link their phone number to their account, starting with Canadian numbers. They can then call their companion directly from their phone or text it, just like a friend. In the app, a button starts a live voice call. All three channels share the same context.
A real-time voice pipeline
Each call follows the pipeline described in our article on how AI voicebots work: streaming speech recognition, response generation by a language model, then speech synthesis in the voice the user picked. The voice, chosen after listening to several options, stays the same everywhere and over time. Tuning latency and turn-taking is what separates a natural exchange from a choppy one, and it sits at the core of our STT, TTS and NLU expertise.
Long-term memory under the user’s control
The companion relies on long-term memory that keeps important information across conversations, whatever the channel. Users can view, edit or delete each memory, and the memory separates real facts from the fictional universe they may build with their companion.
Privacy and operations
Conversations are encrypted in transit and stored in Canada. They are processed by selected AI and telephony providers, and are never sold or used for advertising. Voxivra runs the service continuously, which puts us face to face with production concerns: monitoring, cost per voice minute, call quality and telephony security.
Results
Compagnie is available as a free beta with reasonable usage limits, with paid plans announced. In production, the service demonstrates:
- real-time voice conversations from the app or a plain phone call;
- continuity across text, voice and SMS thanks to shared memory;
- an end-to-end French-language product, hosted in Canada.
What our clients gain from it
Running our own product lets us test in real conditions what we recommend: voice engine selection, latency budgets, memory and personal data handling, operating costs. Those lessons feed directly into our custom AI voicebot projects for businesses.