Research2026

When should a voice agent act?

In LifeMate, a spoken request becomes the right tool calls, one clarifying question, or nothing. Hotword-biased speech recognition keeps names and places right, in English and Vietnamese.

Authors

Summary result

act / ask / quiet

Act

Typed tool calls

Executes logged calls through an MCP sandbox when intent and slots are clear.

Ask

One question

Requests a single clarification when a required contact, date, or slot is ambiguous.

Quiet

No action

Stays silent instead of firing spurious tool calls on non-actionable speech.

The cost of acting at the wrong moment

A voice assistant that acts at the wrong moment is worse than none. Unlike typed chat prompts, live speech includes hesitations, side remarks, partial commands, and mixed English and Vietnamese names.

In LifeMate, we framed voice assistance around an explicit three-way decision policy—act, ask, or stay quiet—rather than forcing every utterance to trigger a tool call.

Keeping names and places accurate across English and Vietnamese

Downstream tool calling fails fast when speech recognition mangles a contact name, calendar title, or place name. LifeMate runs streaming 16 kHz English and Vietnamese speech recognition with hotword biasing so proper nouns survive transcription intact before slot filling and structured LLM tool calling.

Auditable execution boundary

When the policy selects act, the agent fills typed slots (such as contacts and dates) and executes against connected tools—calendar, CRM, or internal APIs—through an MCP sandbox with an action log audit trail.

  • Streaming 16 kHz audio capture with English and Vietnamese ASR + hotword biasing.
  • Explicit act / ask / quiet decision policy before any side effect.
  • Typed, logged tool calls executed inside an MCP sandbox with full auditability.

Have a problem AI should solve?

Tell us what the data looks like and what “working” would mean for you. One of the founders will reply.