Corey James- 8 × 8 sguided enrolment samples, none of them kept
- 3policies: open, shadow and locked
A desk microphone hears everything in the room: the operator, a visitor, the television, loud typing. An assistant that can take actions on a spoken word should know whose word it is. The voice engine has an optional voiceprint filter that recognises the operator's voice and silences every other voice before the engine can hear it, so another speaker cannot interrupt the agent, be transcribed, or trigger a tool.
Enrolment without a recording
Enrolment covers the ways a person actually sits at a desk, through the same echo cancellation and denoising the engine applies to speech, so the profile matches what the engine will hear.
- Embedding of the operator's voice, readable by the operator alonekept
- The audio of the samplesdiscarded
- Profile written only after every sample succeedsatomic
- A failed replacement leaves the old profilesafe
What the filter decides
Incoming audio is scored against the enrolled profile, and the decision is one of three.
The ambiguous case is passed because the filter's job is to keep strangers out.
The cost of silencing the operator is higher than the cost of admitting a doubtful burst.
Switched by asking
The filter has three policies: open, shadow and locked. The operator changes them by speaking. "Listen only to me" locks it; "Let my wife talk to you too" opens it. The model interprets the request, the engine applies it deterministically through a native tool, and the assistant confirms aloud. A command-line tool is the recovery path, and a locked filter with a missing or corrupt profile fails loudly rather than opening silently. In the field, the operator was recognised while walking around the room.
Why not on Teams
The filter runs on the desk microphone only. In a Teams call the engine relies on Microsoft's native voice isolation at the capture device, and the decision is written down with its reasons.
What this demonstrates
I design privacy-sensitive capabilities with their boundaries explicit: enrolled without keeping audio, biased toward hearing the operator, switchable by voice with a loud failure mode, and kept off the surface where it would do harm, with the reasoning written where the next engineer will find it.
The enrolment rule, the ambiguity rule and the Teams decision were mine; AI coding agents, a mix of frontier and open-source models, implemented the filter under my governance system.