Personal AI · Voice interaction

Nyra: deciding when an assistant should speak

A personal AI that can follow a conversation, recognize when help is useful, and offer it at the right moment.

The idea

An assistant should be useful without needing a new prompt for every small thing. With Nyra, I am exploring how a personal AI can follow a conversation, offer a relevant reminder or short cue, and prepare a next step. Knowing when to stay quiet is part of that task.

What I am building

The current voice prototype brings together audio segmentation, speaker verification, model-based intervention decisions, and spoken output. It supports short cues and reversible local action drafts. The broader product direction connects this interaction with personal context and longer-term memory.

How I evaluate it

The project includes recorded-conversation replay and timing instrumentation. The blind replay path receives audio as the conversation unfolds; expected answers are reserved for evaluation afterwards. This makes false interventions, missed opportunities, and response latency visible alongside successful examples.

Where it stands

Nyra is a voice research prototype with a separate product-experience website. The current emphasis is intervention behavior and evaluation. Persistent personal memory and broader integrations remain development directions.

Where it started

I previously built and used Lyra, an end-to-end voice assistant connected to my calendar and tasks. Nyra develops that interest further, with more emphasis on timing, context, and the decision to act.