Your meeting is in French, Hindi, or Spanish, but an English-only transcription model keeps getting it wrong. OGAD (Off Grid AI Desktop) Pro can record the call on your Linux computer and use a multilingual local speech model for the transcript. Prepare the model first, then check the recognized words against the saved recording.
Get OGAD for Linux | 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.
Prepare the Linux beta first
Use the x64 Linux AppImage or deb package from beta 114, with Pro active. This guide covers the Pro code included in that build. Earlier Linux betas did not expose the same feature set.
For screen-based history, prepare the local processing models, review Settings > Setup & health, and allow the capture request your desktop session presents. On Wayland, the desktop portal and your screen or window selection affect what can be captured. Desktop accessibility data depends on what your session and apps expose. Start with a short session before relying on a full day's record.
Choose Resume capture in Replay or Settings > Capture only when you want screen history retained. Check the visible Capturing state. Use Pause capture to stop new screen history. Clipboard, Vault, and imported calendar records are separate workflows; they do not require you to keep screen capture on.
This is useful when you want meeting notes in the language people actually spoke. Transcription is separate from translation: a French transcript is not automatically an English version of the call.
What do you need for multilingual meeting notes?
Use a supported x64 Linux computer, current OGAD Pro, recording permissions, and a downloaded multilingual speech model. A local text model is also needed for generated summaries.
Choose a multilingual Whisper model. An English-only model with .en in its name is not suitable just because the app or computer is set to another language. Complete Pro setup and model downloads while connected.
| Local speech model | Approximate desktop download | When to try it |
|---|---|---|
| Whisper Base, multilingual | 148 MB | A small first model for a short test |
| Whisper Small, multilingual | 488 MB | A comparison if Base misses too many words |
These are download sizes, not total memory requirements. Larger is not a guarantee of accurate recognition. Recording quality, accents, names, and the language mix all affect the result.
How does the meeting choose a language?
The meeting transcription path uses the selected speech model and language setting. In beta 114, Meetings also has a Language control. Select the intended language, or use automatic detection with a multilingual model. Check a short sample before recording a full meeting.
A short name or two-word answer gives detection less information than a complete sentence. For the first check, use a clear sentence with a date or number you can verify.
Switching languages during the same call can be harder than a single-language recording. Do not assume that support for each language guarantees accurate recognition of rapid code-switching or overlapping speakers.
Getting started
- In Models > Transcription, download multilingual Whisper Base or another suitable local model and select Use.
- In Models > Text, select a downloaded local model for the summary.
- Finish the requested microphone and screen/system-audio permissions.
- Join a short agreed test call, open Meetings, and check the recording state. Use Record meeting if needed.
- Have each side say a sentence in the intended language, then select Stop.
- Open the saved meeting's Transcript and compare the words with playback.
Screen content can also be recorded. Make sure participants agree before recording, and keep the visible recording indicator in view. The online meeting platform still carries the call over its network; local transcription means the OGAD speech stage uses your Linux computer.
What should you check in the transcript?
Check names, dates, numbers, and negations first. Those errors can change the meaning even when the rest of the sentence reads well.
For a French call, check that a time such as "dix heures" has not become a different hour. For a Hindi or Spanish call, use a sentence you can verify in that language.
The recording separates your microphone from the other side of the call. It does not guarantee the identity of every remote speaker. Check attribution manually when several people take part.
How do you compare another speech model?
Download the other model first. Open the saved meeting and select it in the Transcription model control, then choose Re-transcribe. Beta 114 keeps speaker-line labels during this processing path; those labels still do not identify each remote participant by name. This replaces the transcript and generated summary, so export a copy first if you want to compare versions.
Compare the same names and numbers in both results. Keep the model that gives a useful transcript on your recording rather than assuming file size alone decides quality.
If no transcript was produced, Retry transcription may be available. Check the local model and the recording before retrying; a missing remote-audio track cannot be recovered by changing the speech model.
Can you translate the transcript afterward?
Yes, as a separate local chat task. Export or copy the checked transcript text and ask a suitable downloaded text model to translate a manageable section. Translation quality and supported languages depend on that model, and a long transcript may need several sections.
A summary is also a separate generated result. It can omit later material from a long call, so keep the full transcript and recording when you need a complete review.
Get OGAD, select a multilingual local speech model, and check a short meeting sample in your language before using it for a full call.
This guide covers OGAD beta 0.0.55-beta.114. The release and build record identify the desktop version and its included Pro code. Setup details follow that code; they are not a device test.
Top comments (0)