OGAD (Off Grid AI Desktop) can infer the spoken language while it transcribes a recording on your device. Download a multilingual speech model, select Auto-detect, and record without an internet connection. This works as part of local speech-to-text. It does not provide a separate language report or a confidence score in the chat composer.
Get OGAM (Off Grid AI Mobile) for your phone | Get OGAD
Auto-detect is useful when you move between languages and do not want to change the setting before each recording. If you already know the language, an explicit choice gives the speech model that information directly.
This guide uses recordings made through the app's chat microphone. The result is an editable transcript. It does not describe a tool that labels every language in an uploaded audio file.
What does automatic language detection actually do?
Auto-detect lets the multilingual speech model infer a language from the audio instead of receiving a fixed language choice. The model then transcribes the speech. You review the words it returns. OGAD's chat workflow does not display a detected-language name or a certainty percentage next to those words.
Keep these tasks separate:
| Task | What you get |
|---|---|
| Automatic language detection during transcription | A transcript produced without selecting a fixed language first |
| Explicit language selection | A transcript produced with your chosen language as a hint |
| Translation | Text in another language, as a separate task |
| A language identification report | A language label or confidence result; not the output of the chat procedure here |
The Whisper speech model supports multilingual recognition and language identification. That capability does not mean every app displays the detected language or exposes all of the model's outputs.
What do you need before going offline?
Install the app for your device and finish the download of a multilingual speech model while connected. Select the local model, give the app microphone access, and choose Auto-detect. Automatic detection runs with the speech model; it does not require a second online language service.
For mobile, use version 0.0.111 or later for the controls below. The published requirements are Android 10 or later with at least 4 GB of RAM, or iPhone 12 or newer on iOS 17 or later.
The desktop controls below are present in release 0.0.49 and later. Use a current desktop release for your operating system, and finish any required runtime and model downloads before disconnecting.
You also need free storage and enough available memory for the model you select. A model's download size is not its complete memory requirement while running.
How do you enable Auto-detect on Android or iPhone?
Select an on-device multilingual speech model, then choose Auto-detect in the transcription language setting. English-only models do not offer automatic selection among multiple languages. The app's language choices follow the active speech model, so change the model first if Auto-detect is missing.
- Open Model Settings > Transcription (Speech to Text) > Transcription model.
- Under On-device models, select Base, 99 languages. Its approximate catalog download is 142 MB.
- Wait for the download and loading to finish.
- Open Chat Settings > SPEECH TO TEXT > Language.
- Select Auto-detect.
- Turn on airplane mode and check that Wi-Fi is off.
- In ordinary Chat mode with a text model, hold the app's microphone, speak, and release it.
- Review the words in the message box before sending or copying them.
Use the microphone inside OGAM, rather than your keyboard's dictation control. Voice conversation mode and models that accept audio directly can follow a different send flow.
If you see a filename ending in .en, such as base.en, that is an English-only model. Choose the multilingual Base entry instead. A multilingual model can still recognize English.
How do you enable Auto-detect on desktop?
In OGAD, select an installed multilingual transcription model and set Spoken language to Auto-detect. Record from the text chat's microphone. The transcript appears in the message draft, where you can check it before sending.
- In Models, download a multilingual transcription model such as Whisper Base.
- Open the model settings panel and choose Transcription.
- Under Current model, select the installed multilingual model.
- Set Spoken language to Auto-detect.
- Disconnect the computer from the internet after setup.
- Open text chat and select Record voice.
- Speak a complete sentence, then select Stop recording.
- Read the transcript in the message draft.
Make sure a remote transcription service is not the active selection. The model settings show which transcription model is used for the next recording.
If Auto-detect is absent, check the model. The desktop language selector offers English only for English-only models. Download and select a multilingual Whisper model for this workflow.
When should you choose a language yourself?
Choose a specific language when you know what the speaker will use, especially for a short note. Use Auto-detect when the language can change between recordings or you do not know it in advance. Treat the transcript as a result to check, not proof that detection was correct.
| Recording | Useful first choice | What to check |
|---|---|---|
| A short note in a known language | Select that language | Names, numbers, and missing words |
| Separate notes in different languages | Auto-detect | Whether each transcript matches its recording |
| A single name or borrowed word | Add context and select a language if known | A name alone may not identify the language |
| One sentence mixing languages | Try an explicit language, then compare with Auto-detect | Words and script around each language change |
| Speech with loud background audio | Improve the recording first | Whether the model is transcribing the intended speaker |
No single recording length guarantees correct detection. Start with a clear, complete sentence that contains more than a greeting or a name. You do not need to create a long recording just to check the setting.
How can you compare Auto-detect with a fixed language?
Use a sentence you understand and can check. Record it once with Auto-detect, then again with the language selected explicitly. Keep the wording and recording conditions similar. Compare the meaning and important details, rather than choosing the result with the most polished punctuation.
For example, include a weekday, a time, and a familiar place in your sentence. Check all of them. If you switch languages often, repeat the check with another language you use.
Recognition can vary with the model, accent, background noise, and the words in the recording.
Does Auto-detect handle several languages in one recording?
It can attempt a mixed-language recording, but the setting does not promise reliable language changes within a sentence. It also does not label each speaker or each language segment in the chat transcript. Check mixed-language results closely and use shorter recordings when they are easier to review.
If you need exact names or technical terms, correct them before using the transcript. If you know the main language and Auto-detect gives poor results, select it explicitly and record again.
Automatic detection is not an instruction to translate everything into English. For translation, first correct the transcript, then give it to a suitable downloaded local chat model with a separate translation request.
Why does it stop working without internet?
Check that the transcription model is local and fully downloaded. A remote speech service still needs its connection. Also distinguish a completed transcript from an AI reply: a local speech model can produce text while a separately selected remote chat model cannot answer offline.
For the complete local workflow, keep both model selections on-device. The initial app, runtime, and model downloads can need internet. The recording and local transcription can then run while disconnected.
Try one complete sentence with Auto-detect while offline. If you know its language, compare it with an explicit selection and keep the setting that gives you a transcript you can verify.
Top comments (0)