You can turn Hindi, Spanish, French, and other supported languages into editable text on your Mac without uploading the audio. OGAD (Off Grid AI Desktop) runs the speech model on your computer. Download a multilingual model, select your spoken language, and record. After setup, transcription works without an internet connection.
| Download OGAD for Mac | Desktop releases |
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@offgridmobileai.co, join our Slack community, or talk to us on Reddit.
This guide covers a short recording through the chat microphone. You speak, stop recording, and check the text before sending it. Use it to draft a message in Spanish, capture an idea in Hindi, or turn a spoken note into text you can copy elsewhere.
The steps use the desktop app. Its controls and model downloads differ from OGAM (Off Grid AI Mobile).
What do you need for local transcription on a Mac?
Use an Apple Silicon Mac, M1 or later, with macOS 13 or later. Install the current stable desktop app and keep enough disk space for your selected speech model. You need internet for setup and model downloads, plus microphone access for recording. The transcription itself can run offline.
The desktop download page lists the supported Mac hardware. These instructions cover the transcription controls available in stable version 0.0.51.
Start with a small model before trying a large one. Free disk space and available memory are different limits: a model can fit on your disk while leaving too little RAM for it to run alongside your other apps.
A chat model is optional for the speech-to-text step. If you want an AI reply, translation, or summary afterward, download and select a local text model too.
Which model should you choose for multiple languages?
Start with Whisper Base in the desktop app’s Models > Transcription category. It is a multilingual model with an approximately 148 MB download. Try a short recording in your language first. If too many words are wrong, compare it with Whisper Small before choosing a larger download.
The desktop catalog includes these multilingual options:
| Model | Approximate download | When to try it |
|---|---|---|
| Whisper Tiny | 78 MB | A small download for a first workflow check |
| Whisper Base | 148 MB | Short notes and a starting point for your voice |
| Whisper Small | 488 MB | A comparison when Base misses words |
| Whisper Medium | 1.53 GB | Longer tests when you have more available memory |
| Whisper Large v3 Turbo | 1.62 GB | Another option to compare with smaller models |
| Whisper Large v3 | 3.10 GB | A larger download when your Mac has memory to spare |
These rounded sizes come from the desktop model catalog. They describe the downloaded files, not peak memory use. They can differ from the mobile app’s model sizes because the apps can offer different files.
The speech engine is whisper.cpp. Whisper models with .en in the filename are English-only. Select the multilingual model when you want another language. Changing your Mac’s system language does not change an English-only model’s capabilities.
The model catalog also contains other speech engines. For the steps below, use Whisper Base so that the same language controls and expected behavior apply.
How do you set the spoken language?
Open the active transcription model’s Settings, then use Transcription > Spoken language. Choose the language you will speak. Auto-detect is available with a multilingual Whisper model. The selected language applies to the next transcription, so check it before you record a note in a different language.
The desktop selector offers English, French, German, Italian, Spanish, Korean, Japanese, Mandarin, Hindi, and Portuguese, plus Auto-detect. This list describes the app’s explicit language choices. It is not a claim that the model has identical accuracy in each language.
Choose a specific language for your first test. Very short phrases give automatic detection little context. If your name or a two-word note comes back in the wrong language, select the language directly and record a full sentence.
The Voice tab controls spoken output. Use the Transcription tab to control recognition of your speech. Changing the voice used for AI replies does not set the language of your microphone recording.
How do you transcribe speech without uploading it?
Select a downloaded local speech model, set the spoken language, then record through the microphone beside the text message box. Click again to stop. OGAD converts the recording into text and puts it in the message box. You can correct or copy the text before sending it.
1. Download the multilingual model
Open Models and select Transcription. Find Whisper Base and click Download. Wait until the download finishes.
2. Make the model active
Select the downloaded model and use Use this model. Open Settings for the active model. Under Transcription, check Current model.
Make sure a local Whisper model is selected. A remote transcription model changes where the audio is processed.
3. Choose your language
In the same settings panel, set Spoken language to the language you will speak. Close the panel and return to chat.
4. Record a short sentence
Keep the text message box visible. Click its microphone control, labeled Record voice. Allow microphone access if macOS asks for it.
Speak one sentence, then click Stop recording. Wait for transcription to finish.
Use the microphone beside the text composer for this test. The separate Voice conversation mode can send a transcribed turn to the AI. The text composer lets you review the words first.
5. Check the result
The transcript appears in the message box. If you already typed something, the recognized text is added after it.
Correct names, dates, and numbers. You can copy the text into your notes without sending it to a chat model. If you send it for an AI reply, the selected text model handles that next step.
How can you check that transcription works offline?
Finish the download, select the local speech model, and disconnect your Mac from the internet. Turn off Wi-Fi and disconnect Ethernet or any other active network connection. Then record a new sentence through Record voice. The expected result is editable text in the message box.
Use a sentence you can check without guessing. For example, say the equivalent of “The workshop starts on Friday at ten” in your chosen language. Compare the day and number with what you said.
This checks the transcription workflow on your Mac. It does not measure accuracy across languages or establish a processing-speed benchmark.
If the transcript appears but an AI reply fails, check the text model separately. A local speech model can transcribe offline while a remote chat model still needs a connection.
What should you check when the transcript is wrong or missing?
| Problem | What to check | Next action |
|---|---|---|
| Only English appears in Spoken language | The selected model or engine may be English-only | Select multilingual Whisper Base, then reopen the language setting. |
| The output uses the wrong language | Auto-detect or a previous language choice may be active | Set Spoken language explicitly and record a full sentence. |
| The microphone cannot start | macOS may have denied access | Use the app’s Open System Settings button and allow microphone access. |
| No transcription model is found | A download may be missing or incomplete | Finish downloading a model in Models and make it active. |
| A large model fails to run | Other apps or loaded models may use the available memory | Close memory-heavy apps or try Base. |
| The text is sent before you can edit it | Voice conversation mode may be active | Return to the text composer and use Record voice there. |
| It works online but fails offline | A remote transcription model may be selected | Select the downloaded local model and repeat the disconnected test. |
Reduce background noise and keep a steady distance from the microphone. Try separate recordings for separate languages before testing speech that switches languages within one sentence. Review each transcript even when its punctuation looks correct.
Does the local workflow keep audio off cloud servers?
With a local transcription model selected, OGAD runs recognition on your Mac. The microphone workflow passes the recording to the local speech engine. It does not require a cloud speech service.
A remote transcription selection changes that path. Sending the finished text to a remote chat model is also a separate network action. Keep both selections local if you want recognition and AI replies to run on your Mac.
Local processing does not mean the transcript cannot leave later. You choose where to paste, send, or share it.
Mac transcription FAQ
Do I need Pro for this microphone workflow?
No. Speech-to-text in the desktop chat interface is included in the free app. System-wide dictation and the meeting recorder are separate Pro workflows. The desktop product page distinguishes the free studio from Pro features.
Can I turn Hindi speech directly into an English note?
The steps above transcribe the spoken language. For an English note, review the Hindi transcript and ask a downloaded local text model to translate it. Check names and numbers again after translation.
Will a larger chat model improve speech recognition?
Speech recognition uses the transcription model. Changing only the chat model does not change that step. Compare speech models using similar recordings when you want to improve recognition.
Can I use this procedure for a saved meeting recording?
This procedure covers new audio from the chat microphone. A saved recording or a full meeting uses a different input workflow. Do not play a meeting through your speakers and assume the microphone captures both speakers clearly.
Download OGAD, choose Whisper Base, and try one sentence in your language with the internet disconnected.