Models and Computing Power: What Your Computer Can Do
memozero computes on your own device. Your audio file never leaves the device, and no data center works for you. That is why your hardware decides the speed of your transcription. This article shows how memozero detects your equipment, which models follow from that, and where you can look up the result.
How memozero detects your hardware
At every start, memozero checks which compute acceleration is available. The status line “Detecting hardware” runs on the start screen while it does so. The check looks in this order:
| Equipment | What memozero uses |
|---|---|
| Mac with an Apple Silicon chip | the graphics unit through Metal |
| device with an NVIDIA graphics card | the graphics card through CUDA |
| Windows device with AMD or Intel graphics | the graphics card through Vulkan |
| everything else | the processor |
The result controls which models memozero offers and which path it computes on. You set nothing for this. memozero has no list that lets you pick a compute device. If you add a graphics card later or install a new driver, memozero detects it automatically at the next start.
The detected hardware also decides about speaker detection. On older devices with weak hardware, the counter and the switch can therefore be locked under + New Transcription. You still get a complete transcript; all words then carry an unknown speaker. You can still assign the speaker by hand, however.
Which models are available to you
Under + New Transcription, after you select the file to transcribe, you choose between the different AI models. Select the model mainly according to the language of your recording. The following models are available to you for the following use cases:
| Label in the app | What the app says about it |
|---|---|
| Whisper 3 Turbo – German | Specially trained for German, the most accurate for German recordings, and fast. |
| Whisper 3 Turbo – Multilingual | Faster processing for recordings in other languages or with several languages (for example when an interpreter is used). |
| Whisper 3 – Multilingual | For recordings in other languages or with several languages (for example when an interpreter is used), at the best quality. For German recordings, the German model is more accurate. |
On devices without graphics acceleration, the same models also appear in the list with the addition (CPU). If you have a supported graphics card or an Apple Silicon chip, you do not see these lines. A model that has not been downloaded yet carries the addition (Download).
The list depends on the device you use. Not every user therefore sees all lines. On Macs with less than 16 GB of memory, memozero only offers a leaner version of the German Turbo model. If a colleague has a different selection in front of them than you do, that is not an error.
The first download
The models are not in the installer. memozero downloads them from a server as soon as they are needed. Which models these are depends on the detected hardware. memozero always downloads speaker detection with them.
- The download starts during setup already, while you type in your activation code.
- memozero downloads the transcription service first, then the models.
- The progress appears as a banner or card in the sidebar, with percent, speed, downloaded amount, and estimated time left.
- You keep working in the meantime; the download does not block the app.
- If you close the banner, the download still continues.
If a model is missing for a job, that holds up only this one transcription. The app stays usable, and your other transcripts still open.
What determines the duration and quality of the transcription
The speed of a transcription varies in particular depending on:
- the length of the recording. Two hours take longer than ten minutes.
- the quality of the recording. Clear, loud voices and little background noise are faster and more accurate to process than a noisy recording with many speakers that is interrupted or overlaid by background noise or interjections.
- the accelerator. A device with Metal, CUDA, or Vulkan computes faster than a device that only uses the processor.
- the memory. If too little is free, memozero shows the dialog Low memory available before the start, with the option to continue via Start anyway.
- the selected model. Turbo is the fast mode, “Whisper 3” the thorough one.
- the queue. Only one transcription runs at a time. Other jobs wait in the order they came in.
Keep memozero open until the transcription is finished. Closing the app ends the running transcription.
Submit very long files in the evening and leave memozero open overnight. The next morning, the job is set to Review, and the editor waits for you.
No quota
memozero does not sell minutes. There is no minute quota, no upper limit for the number of transcriptions, and no limit for the length of a recording. The number of waiting jobs is unlimited as well.
Longer recordings (with good audio quality) make it easier for speaker detection to assign speakers, because more material per speaker is available.
Only your patience and your hardware set limits. For the size of the file to be transcribed, a hard limit of 50 GB applies. From 2 GB on, the app warns you, and on devices with 8 GB of memory it warns from 1 GB on. The text reads: “Large files are supported, but offline processing can take a long time. Keep memozero open until the transcription finishes.”
Do you want to find out what your device can do?
- Click the email address in the footer of the sidebar. The account settings open.
- Go to the tab System.
- Click View System Info.
The window shows the version of the app, the operating system, the detected compute device, and the state of CUDA, Metal, MLX, and FFmpeg. Below that, you find the installed models and the state of the two services.
That answers two questions in a few seconds: does memozero really use the graphics card on this device, and is the model you want already on the disk?
For a support request, for example after a failed transcription, export your support package on the System tab under Diagnostics, save it, and send it to info@memozero.io. The support package contains no personal data such as your email address. File names and file paths are redacted in it; only the file extension remains.