You can speak a question in the Off Grid AI Chrome extension and hear the answer. OGAD (Off Grid AI Desktop) handles transcription and speech output with the models selected on the desktop. Use local speech and text models to keep those AI operations on your computer.
The browser microphone produces a draft you can check before sending. This is useful for a longer question about a page, especially when typing would interrupt the work you are doing.
Download OGAD | Desktop beta 122 release notes
What would you like to do with Off Grid AI?
Have a feature or use case you would like us to support? Tell us what you want to do and which device you use.
Write to support@getoffgridai.co, join our Slack community, or talk to us on Reddit.
Prepare all three parts of the conversation
A spoken exchange uses more than a text model. Speech recognition turns your recording into text. The chat model answers that text. Speech output reads the reply aloud.
| Part | What to confirm in OGAD |
|---|---|
| Your spoken question | A local transcription model is ready |
| The written answer | A local text model is loaded |
| The spoken reply | A supported local voice model and voice are selected |
Check these parts separately if voice has not worked before. A successful chat reply does not prove that the microphone or speech model is ready. A recorded clip does not prove that its transcript is accurate.
The available voice choices can differ by desktop platform and installed package. Select a voice that the app offers on your computer rather than assuming that every voice shown elsewhere is installed locally.
Connect Chrome and allow the microphone
Get the early-access extension and pair it through Settings > Desktop. Compare the six words with the request in OGAD and approve the matching pair. Keep the app running while you use speech.
Start the microphone from the composer. Chrome's side panel can need a separate extension tab for the first microphone permission. If that tab opens, allow access there, return to the side panel, and try again.
If no microphone is found, check the device selected by the operating system and browser. A headset that disconnected during a call may leave a different input selected. Use a short recording to confirm that the correct microphone is active.
Dictate a useful page question
Open the page you want to discuss. Start recording, speak one complete question, then stop. The extension sends the clip to the desktop transcription service and puts the returned text in the composer.
For example:
Explain the three export options on this page. Tell me which one keeps the original page size. Use the wording on the page, and do not change any settings.
Read the draft before sending it. Correct product names, numbers, and words such as “not”. One missed negative can reverse the task. If the transcript is mostly wrong, make a shorter recording in a quieter setting instead of editing a long paragraph.
The first success is a correct transcript that you deliberately send. You do not need a hands-free setup to get value from voice input.
Hear the reply and keep the text available
Use Read aloud on the answer. The desktop generates speech and the browser plays it. Use Stop if you need to end playback. Keep the written answer visible so you can inspect a command, filename, or number that is difficult to understand by ear.
Ask for a short answer when you intend to listen:
Give me the recommended option in one sentence, then the two checks I should make before using it.
A long table can be useful on screen and poor as speech. Ask for a brief spoken summary while retaining the detailed text for reference. The model should preserve the meaning, not omit a condition that changes the instruction.
Use voice without losing control of the task
Dictating in the browser does not require an always-listening wake word. The recording control makes the input step visible. Stop or cancel the recording when you are done.
Page context can accompany a browser question. That is useful when your request says “this option”, but you should still name the relevant section when several options are visible. Check the transcript and page together before sending a request that could cause an action.
Speech output can be heard by other people in the room. Use headphones for private work. Local inference changes where the models run; it does not make an audible reply private to the person who asked.
Fix one stage at a time
If nothing is recorded, check browser and operating-system microphone permissions. If recording works but no text returns, check the desktop transcription model and connection. If the answer arrives but no voice plays, check Read aloud, the voice model, and the browser output volume.
Downloads need internet during setup. With local models ready, the speech and text processing can run without a cloud AI service. A live website or a desktop reached across the internet still has its own connection requirement.
Set up OGAD and try a ten-second question. Review its transcript, send it, then use Read aloud on the answer.
Top comments (0)