The problem
A telephone conversation is a streaming interaction with awkward timing. Recognition, model latency and speech playback all share a caller’s patience budget.
The architecture
RizoReception connects a PBX audio path to speech recognition, a local language model and speech synthesis. Transfer policy is a separate boundary. A plausible sentence from a model is not authority to send a call anywhere it likes.
The hard part
A good response in a browser does not prove a good response over a telephone. Audio quality, pauses, interruptions and the actual transfer path all need acceptance evidence.
Documented milestones
Earlier work recorded human-call transfer acceptance. A later voice activation had its own automated checks and outstanding listening and transfer confirmation. Those records must remain distinct: a previous voice’s acceptance cannot qualify a replacement automatically.
Current limits
The project is presented as experimental. No caller data, private destinations, recordings or transcripts are published here. The dual-card tower photograph is historical build imagery, not the current AI guest specification.