Voice-controlled desktop automation

Speak a task. Keep control.

Heliox can turn opt-in voice input into the same typed, permission-checked action flow used by the desktop app—so speaking does not bypass planning, approvals, or verification.

Opt-in microphoneContinuous listening availableSame approval policy as text

Implemented; hardware-dependent

What Heliox does

Recognizes a wake phrase or an explicit listening session and converts speech to a command.
Routes the command through Heliox's declared action registry and specialist agents.
Keeps listening after a completed command when continuous mode remains enabled.
Can speak status, interruption prompts, and results through the configured local or operating-system voice.
Runs Pocket or Kokoro inference in a bounded worker process that is reused for a short speech burst, then exits so heavy model libraries are not retained by the control daemon.

Typical flow

Enable voice and select a microphone.
Say the wake phrase and a concrete request.
Review any required approval.
Hear and read the verified result.

Safety boundary

Voice is an input channel, not extra authority. A spoken request cannot elevate permissions, accept its own approval, or remove a deterministic safety warning.

Known limits

Where this can fail

Recognition quality and response time depend on the microphone, room, platform speech service, selected model, and requested action. Neural voice cold-start time depends on the host; the worker temporarily uses model memory while active. Physical capture and audible output must be tested on each device.

Inspect the evidence

Read generated coverage, measured latency, hardware boundaries, and known limitations.

Open proof center