Speak a message

3 minute read · Chats & Local Agents

Speak the next change when it is easier than typing. Ordinary dictation puts recognized text in the draft so you can review it before sending.

• Set up your microphone
• Try a spoken conversation
• Dictate a draft first
• Use spoken replies deliberately
• Solve common audio problems

Set up your microphone#

Open Voice settings, choose the input device and language, and grant the operating system’s microphone permission. First use downloads the local Whisper model when needed. If nothing is recognized, check the selected microphone, permission and model state before retrying.

Recognition runs locally. Audio is buffered in memory rather than uploaded as a speech request or saved as an audio file. The recognized text enters your ordinary draft; once submitted, it is part of the interaction with your selected provider.

Try a spoken conversation#

Enable spoken replies when you want that behavior. After a pause, the voice conversation can submit your message, read the final response and resume listening. Capture pauses during speech playback to avoid hearing itself.

Provider questions and approvals still need your answer. Changing chat, account, model or access mode ends the voice session; hiding the app pauses capture. Available system voices depend on your OS, and an unavailable voice should be reported rather than silently using another language.

Continue with drafts and approvals.

Dictate a draft first#

Choose the intended chat, model and access mode before starting voice input. Check the microphone, recognition language and sensitivity in Voice settings. Press the microphone and speak a short request. Ordinary dictation inserts recognized text into the draft; review and correct it before sending.

Recognition runs locally using a lazily downloaded Whisper model in an Electron utility process. The audio is buffered in memory rather than uploaded as a speech request or saved as an audio file. Once you submit the recognized text, it becomes an ordinary provider interaction with that provider's usual data and usage boundary.

Use spoken replies deliberately#

Spoken replies are opt-in. In a voice conversation, a pause can submit your request to the chosen agent, read its final response and resume listening. Capture stops during speech playback to avoid recording the app's own voice. A question or approval still needs your answer; voice mode does not grant extra permissions.

Changing chat, account, model or access mode ends the voice session. Hiding the app pauses capture. Before moving to sensitive work, confirm that the microphone is no longer listening rather than assuming an old chat's voice state applies everywhere.

Solve common audio problems#

If nothing is recognized, check OS microphone permission, the selected input and whether the local model download completed. If the words are wrong, verify the recognition language and test a short sentence in a quiet environment. If a spoken reply is silent, check the output device and installed system voice.

System speech voices depend on the OS. An unavailable voice should be reported rather than silently switching languages. Local recognition does not consume provider tokens until text is submitted for agent work; Vortex does not include the external provider's allowance. Model usage explains the distinction.