Guide
Transcribing calls on your own computer: what you need
Updated
The short answer
Odrivyn can transcribe calls on your computer instead of in the cloud, so the call's audio never leaves it and no transcription minutes are used. It is a free download in Settings. You need Windows 10 or 11 (64-bit) with a processor from about 2013 or later, 4 GB of memory for the smallest option and 8 GB or more for the larger ones, and room on your disk for the download.
What does on-device transcription do?
Normally Odrivyn sends the call's audio to a transcription service and gets the words back. With on-device transcription, a model on your computer turns speech into text instead.
- The audio never leaves your computer.
- It works without an internet connection.
- It does not use your transcription minutes.
You can use it all the time, or choose "Cloud, on-device if offline": the cloud first, and your computer when the cloud is out of reach or your minutes run out.
Which option should I choose?
There are four. The bigger ones are more accurate and need a stronger computer. If you are not sure, start with Base; you can add another later and switch in one click.
| Option | Download | Memory while running | Best for | Languages |
|---|---|---|---|---|
| Base | 148 MB | about 0.4 GB | Any computer; clear audio | 99 |
| Small | 190 MB | up to about 0.9 GB | A good balance of speed and accuracy | 99 |
| Large v3 Turbo | 574 MB | about 1 GB (estimate) | The most accurate; needs 8 or more processor threads | 99 |
| Parakeet | 670 MB | at least 2 GB | Fast and accurate in European languages | 25 European languages |
Memory figures are the published ones for these models; Turbo has none, so it is our estimate. Parakeet's languages: Bulgarian, Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hungarian, Italian, Latvian, Lithuanian, Maltese, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Ukrainian.
What does my computer need?
- Windows 10 or 11, 64-bit. Windows on ARM and Mac are not supported yet; calls there are transcribed in the cloud.
- A processor with AVX2: Intel from 2013 on, AMD from 2015 on, and every AMD Ryzen. If yours is older, Settings says so and calls stay in the cloud.
- Memory: 4 GB is enough for Base. For Large v3 Turbo and Parakeet, 8 GB or more; 16 GB is comfortable, because your call app and browser need memory too.
- Processor threads: 4 or more; 8 or more for Large v3 Turbo.
- Disk space: the download plus about half a gigabyte to spare.
- Nothing else to install. The engine comes with the files it needs, and the download happens once.
Odrivyn checks all of this for you. Settings shows what your computer has, and marks an option that may not suit it before you download.
How do I set it up?
- Open Settings and find Transcription.
- Press Download next to an option. The engine (17 MB) downloads first, then the model.
- When it says Ready, choose "On this computer" or "Cloud, on-device if offline".
To try a different option, download it and press Use. Remove frees the space again.
What if something goes wrong?
- "The processor is too old": your computer lacks AVX2. Calls keep working in the cloud.
- A warning next to an option: it may fall behind on this computer. A smaller option usually keeps up.
- The download failed or did not match: press Download again. It carries on from where it stopped, and every file is checked before it is used.
- Not enough disk space: free some space, then download again.
- Your antivirus removed the engine: allow Odrivyn, then press Download again.
- The words appear late: choose a smaller option, close heavy apps, and plug a laptop in; on battery, Windows slows the processor down.
- Words in the wrong language: set your call language in Settings. Parakeet only knows the 25 languages above.
Can the answers be written on my computer too?
Yes, with Ollama, a free app that runs AI models on your computer. Install it, download a model (for example, in a terminal: ollama pull qwen3), then in Settings, under Your own model, choose Ollama and press Test.
- Odrivyn turns a model's "thinking" off where the model allows it, so answers start quickly. A model that cannot switch it off (gpt-oss) is asked to keep it short, and its thinking is never shown.
- Odrivyn asks Ollama for enough room to read the whole call. If a model cannot read that much, Odrivyn says so instead of answering from part of it.
- Choose a model whose download is well under your computer's memory. Test loads it, so press Test before a call and the first answer will not wait.
- Models made only for search (embedding models) cannot write answers; Odrivyn tells you if you pick one.
- A model that cannot see images answers without the screenshot, and Odrivyn notes that.
Frequently asked questions
Is on-device transcription free?
Yes. The download is free, and transcribing on your computer does not use your transcription minutes.
Does anything leave my computer?
Not the call's audio. It is turned into text on your computer. Answers still come from the AI you have chosen, unless that is a model running in Ollama on your computer.
Is it as accurate as the cloud?
The larger options come close in clear audio. With background noise or several people talking over each other, the cloud is usually more accurate.
Can I use it on a Mac?
Not yet. On a Mac, calls are transcribed in the cloud.