You can run an existing recording through another local speech model in OGAD (Off Grid AI Desktop). Keep the original audio, save a copy of the first transcript, select a different model, and use Re-transcribe. Compare the new words with the recording before deciding that the result is more accurate.
| Download OGAD | Current 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.
A second pass is useful when the first transcript misses names, confuses numbers, or loses technical terms. A larger model is a candidate to test. It does not guarantee a better transcript, and it cannot recover words that the microphone never captured.
Which re-transcription workflow should you use?
Use the saved recording’s Re-transcribe action if it is in the Pro Voice or Meetings library. For an audio file attached in ordinary chat, select another transcription model and attach the original audio again. These are different workflows, with different effects on the existing transcript.
| Where the recording is | How to run another pass | What happens |
|---|---|---|
| Pro Voice library | Select another speech model, then Re-transcribe | The existing recording’s transcript is replaced |
| Pro Meetings | Choose a model in the meeting view, then Re-transcribe | The meeting is transcribed and summarized again |
| Ordinary chat attachment | Change the transcription model, then attach the original audio in a new chat | A new attachment gets a new transcript |
The current desktop product page lists Pro as available on macOS. The core audio-attachment route is separate and is available in the desktop app without the Pro recording library. The workflows below are present in release 0.0.49 and later.
What must you keep before trying a second model?
Keep the source audio and a separate copy of the first transcript. A transcript alone does not let a speech model listen again. The Voice library’s re-transcription action updates the same recording, so copy the current text before replacing it if you want a comparison.
Before you start:
- Check that the recording still plays.
- Copy or export the current transcript to a separate note.
- Write down which model produced it.
- Download the replacement model and check that it supports the recording’s language.
In Voice settings > History, Keep limits the number of recordings retained. Auto-delete removes recordings older than the selected period. When the app removes a Voice recording, it also removes its stored audio. Changing the retention setting afterward does not restore a file already deleted.
Keep a separate original file if you expect to compare recordings over time.
How do you re-transcribe a recording in the Voice library?
Open Voice, choose another installed speech model in Voice settings, then use Re-transcribe on the saved recording. The app uses the retained media and updates that recording’s transcript. Wait for processing to finish before comparing it with your saved first version.
1. Save the first transcript
Find the recording in Voice. You can use Search transcripts to locate it. Select Copy transcript on its card and paste the text into a note outside the recording library.
Use the recording’s playback control to check that the audio is available.
2. Select another local model
Open Voice settings using the gear control. In Transcription, choose the appropriate speech engine. The model list below it shows the models for that engine.
For a comparison within multilingual speech recognition, you can try another multilingual Whisper model. Select Download if it is not installed. When ready, select Use and check that the model is marked Active.
Keep the comparison specific. For example, compare Base with Small on the same recording. If you change both the model and the recording, it is harder to tell what caused a difference.
3. Run the second pass
Return to the recording card and select Re-transcribe. The action is unavailable while that recording is already being processed or if it has no stored audio path.
The existing card receives the new result. This action does not create a second recording for your comparison, which is why saving the first transcript matters.
4. Check the changed words
Compare both versions while listening to the same passages. Keep the version that matches the audio more closely. Correct remaining errors before using the text in notes, summaries, or messages.
How do you re-transcribe a saved meeting?
Open the meeting in Meetings. Use its Transcription model selector to choose an installed model, then select Re-transcribe. The saved meeting media must still be available. The app runs transcription and summary processing again for the same meeting.
Save any transcript or summary you want to retain before starting. A new transcript can change the summary too, so review both after processing finishes.
If the recording is unavailable or incomplete, choosing a larger model will not fix the missing media. Use the original file if you kept another copy.
For an offline run, the models used for transcription and summary generation must already be downloaded and selected locally. A local speech model does not make a separately selected remote summary model local.
How do you retry an ordinary audio attachment without Pro?
Select another local transcription model, then attach the original audio file in a new chat. OGAD extracts a fresh transcript using the active transcription model. Open the attachment preview to compare the text with the first result.
- Download the alternative speech model from Models.
- Open model settings, select Transcription, and change Current model.
- Set Spoken language if you know it, or use Auto-detect with a multilingual model.
- In a new chat, select + > Attach files and choose the original audio.
- Wait for processing to finish, then click the text preview to expand it.
Use an audio file such as MP3, WAV, or M4A for this route. The ordinary chat video attachment path samples visual frames; it does not transcribe the complete soundtrack. Export the audio first if the original is a video.
How can you tell whether the new model is more accurate?
Judge the transcript against the recording, not against how smooth the sentences sound. Check the same difficult passages in both versions. Count errors that matter to your task, such as a wrong name or missing negative, rather than treating punctuation differences as equal to a changed fact.
| Detail | What to compare |
|---|---|
| Names and technical terms | Correct spelling and whether the word was actually spoken |
| Numbers and dates | Amounts, units, times, and order |
| Negatives and conditions | Words such as “not,” “unless,” and “only” |
| Quiet speech or interruptions | Missing words and invented text |
| Multiple languages | Whether each passage keeps the intended words and meaning |
Check more than one passage. A model can fix a name near the start and make a new error later.
For a fair first comparison, keep the same audio and change only the model. If you later test language settings or custom vocabulary, record that as another change. There is no single model that this guide can promise will win on every voice and recording.
What if re-transcription fails?
Check that the original media exists, the model is installed, and enough memory is available. A download on disk still needs memory to run. If the larger model cannot load, unload unused models or choose a smaller one.
| Problem | Next check |
|---|---|
| Re-transcribe is disabled | Wait for the current run, or check whether the recording still has stored audio. |
| The recording cannot play | Find your original copy; retention or deletion may have removed the library copy. |
| No transcription model is installed | Finish downloading a speech model and make it active. |
| The result is still wrong | Check language support and listen for noise or unclear speech. |
| The first transcript is gone | Re-transcription updates the existing item. Use the copy you saved before the run. |
Download the comparison model before going offline. Then save the first transcript, run one recording again, and check the parts that matter before processing the rest of your library.