When you talk to XNeuronal, everything seems instantaneous. Behind the scenes, three distinct layers work in sequence to transform your voice into a response, and it's an architectural choice we fully embrace.
Three Layers: Hear, Think, Speak
The journey of a sentence always follows the same path. First, transcription (STT): audio becomes text. Then, reasoning: a language model understands the intent, organizes the neurons, and decides what to say. Finally, synthesis (TTS): text becomes voice again. Each layer is an independent service, replaceable without affecting the others. This modularity isn't a luxury: it allows us to switch providers when a better one emerges, without rewriting the product.
Always a Safety Net
An external service can be slow, overloaded, or fail. So, we designed each layer in a cascade: if the primary engine fails, a backup engine takes over without the user noticing. Transcription switches from one provider to another, synthesis falls back to a voice of the same gender. The goal is simple: a provider outage should never become an assistant outage.
Why Push-to-Talk, and Not Interruption
We could have aimed for continuous conversation from the start, where the AI interrupts you and responds before you finish your sentence. We chose the opposite: a simple press to speak, a release to listen. This isn't a lack of ambition, it's a calculation. The real-time voice market changes every month; pre-coding full-duplex for one provider today means writing dead code in six months. Better to have a lean and solid foundation, ready to adopt this choice at the right time.
The Latency We Accept, The Latency We Refuse
That's the whole challenge: what waiting time is tolerable. One to two seconds for a thoughtful response is acceptable, it's even reassuring, it signals that the assistant is thinking. What we refuse is forced waiting: a silent failure, a voice that never comes. The pipeline's simplicity is precisely what allows us to control this boundary.