The Whisper and cleanup models YapToText runs on your Mac
By Ryleigh Newman · Published · Updated
YapToText runs Whisper models locally on your Mac: Whisper Large v3 Turbo (Q5), 574 MB, turns your voice into words, and Phi-3.5 Mini Instruct, shown as 2.3 GB, cleans those words up when a mode asks for it. Both ship inside the App Store build, so there is nothing to download and nothing leaves the machine. The AI Models page is where you see them, swap either one, or add a model file of your own.
I picked the 5-bit build of Whisper because, on my own dictations, it heard almost exactly what the full model heard, with a third of the memory. It is the standard whisper.cpp file, not a custom-trained one. For Whisper on a Mac in general, the Whisper Mac app guide has the bigger picture.
This app is completely free. You don't have to make an account. There are no subscriptions, no difficult downloads, and nothing is locked. Made for Apple silicon Macs. Source on GitHub.
On this page: 9 sections
Two models, two jobs
- The dictation model. Whisper Large v3 Turbo (Q5), the file ggml-large-v3-turbo-q5_0.bin: 574 MB, multilingual, MIT licensed, marked Recommended. It has been the default since 1.3.1, the release whose notes say dictation became over twice as fast and the app about a gigabyte smaller.
- The cleanup model. Phi-3.5 Mini Instruct, a llama.cpp GGUF file the app shows as 2.3 GB, multilingual, MIT licensed, marked Recommended. Clean Up, Note, Email, Message, Code Comment, Auto mode's formatting and Quick Edit use it. Raw Transcription never does.
- Apple Intelligence. On macOS 26, where it is available, it can take the cleanup job if you choose it. It is not the default, and before macOS 26 the choice is hidden.
Under the hood is a copy of whisper.cpp with llama.cpp beside it, built with Metal and Accelerate. Every mode's editor has its own Speech model and Cleanup model choices, and a mode's choice wins over the defaults. Whisper hears you. The cleanup model only rewrites, and only when the mode asks.
Pick another model from the library
The AI Models page opens with Your defaults, the two models doing the work today, each under a plain label for its job, with the Language picker under the dictation model. Below that are the Dictation Model Library and the Cleanup Model Library.
Step 1: Download a model
- You: In one of the libraries, press the download button at the end of a row.
- YapToText: Fetches the file over HTTPS from Hugging Face, with a progress bar you can cancel.
- Check: The row says Installed. If the download failed, it shows Retry.
Step 2: Select it
- You: Click the circle at the start of the row.
- YapToText: Uses that model from your next dictation.
- Check: The circle is filled. A model can only be selected once it is fully on disk.
| Model | Job | Size the app shows | License | Note |
|---|---|---|---|---|
| Whisper Tiny (English) | Dictation | 75 MB | MIT | English only |
| Whisper Base (English) | Dictation | 142 MB | MIT | English only |
| Whisper Small (English) | Dictation | 466 MB | MIT | English only |
| Whisper Base | Dictation | 142 MB | MIT | |
| Whisper Small | Dictation | 466 MB | MIT | |
| Whisper Medium | Dictation | 1.5 GB | MIT | |
| Whisper Large v3 Turbo (Q5) | Dictation | 574 MB | MIT | Bundled, the default |
| Whisper Large v3 Turbo (Q8) | Dictation | 834 MB | MIT | |
| Whisper Large v3 Turbo | Dictation | 1.5 GB | MIT | |
| Whisper Large v3 (Q5) | Dictation | 1.1 GB | MIT | |
| Whisper Large v3 | Dictation | 3.0 GB | MIT | The largest and slowest |
| Phi-3.5 Mini Instruct | Cleanup | 2.3 GB | MIT | Bundled, the default |
| Llama 3.2 1B Instruct | Cleanup | 808 MB | Llama 3.2 Community | |
| Llama 3.2 3B Instruct | Cleanup | 2.0 GB | Llama 3.2 Community | |
| Qwen 2.5 3B Instruct | Cleanup | 1.9 GB | Qwen Research | Described as especially for non-English dictation |
| Qwen 2.5 1.5B Instruct | Cleanup | 990 MB | Apache 2.0 | |
| Gemma 2 2B | Cleanup | 1.7 GB | Gemma | |
| Apple Intelligence | Cleanup | Built in | Apple system | macOS 26 only |
Installed models have a menu with Show in Finder and Remove download. The two bundled models say Included and cannot be removed. If the speech model you picked is not on disk yet, another downloaded Whisper model stands in, so dictation keeps working while it downloads. Bigger models hear better. Smaller ones are lighter on battery and memory.
For what each Whisper size costs in memory, the local Whisper guide lists whisper.cpp's own figures.
Read the stars
Every row carries five stars. Click them and two plain ratings appear, one for accuracy and one for performance, each high, good or low, plus Recommended on the pick for most people. Treat them as a guide for choosing between models. The stars help you choose. Your own dictations are the real test.
To see how a model does on your own Mac, open any dictation in History and click Show pipeline details: the stop-to-text time is split into transcription, AI cleanup and delivery.
Add your own model
- You: Under Your own models, press Add Model and pick a whisper.cpp speech model (.bin) or a llama.cpp cleanup model (.gguf).
- YapToText: Reads the start of the file to tell which kind it is, copies it into its Models folder and lists it everywhere a model can be chosen, the Energy page included.
- Check: The model appears under Your own models. A file that is neither kind is rejected with a message saying so.
A GGUF file can still use an architecture the built-in engine does not support, and its row says so: if it fails to load, that is why. Removing one of your own models removes its copied file too. Your own model needs nothing configured: the file itself tells the app what it is.
Where the models live
The Storage card says it plainly: models are stored locally on your Mac. Reveal in Finder opens the Models folder. Downloaded models and your own live in the app's data container, not inside the app, so they survive updates. The two bundled models ship inside the app. Erase All Data, in Settings, Advanced, keeps downloaded models unless you choose to erase them too. Nothing in the Models folder is ever uploaded.
Two settings on the Energy page decide how the models use your Mac: Keep models in memory, 2 minutes after use by default, and Switch models with the power source, off by default, which gives each model a plugged-in and an on-battery choice.
Limits
- The Homebrew build. It ships without models. Download Whisper Large v3 Turbo (Q5) from the Dictation Model Library before your first dictation, and Phi-3.5 Mini Instruct for Auto mode and the other AI modes. Until a speech model is installed, YapToText says no speech model is installed.
- Languages. The Language picker offers Auto-detect language plus 29 languages, and the default is your Mac's system language. The three English models hear English only. Dictating in other languages has more.
- Bigger is slower. Whisper Large v3 is the largest and slowest model in the library.
- No accuracy numbers. I do not publish accuracy percentages for any model. The stars and your own History are the guide.
- Apple silicon. YapToText needs an Apple silicon Mac with macOS 14 or later, and nothing is claimed for Intel Macs.
File transcription uses the same speech model, so the guide to transcribing an audio file on a Mac applies here too, and offline dictation covers working with Wi-Fi off.
If a model misbehaves
- Wait out the first dictation. The model loads once per launch, and again after it has been unloaded to save memory. The dictations after that are quick.
- Retry a stuck download. Files come straight from Hugging Face. YapToText has no server of its own. Press Retry, or check your connection.
- Let it finish. A model that is still downloading cannot be selected yet.
- Check the file type. Your own model must be a whisper.cpp .bin or a llama.cpp .gguf. A GGUF that fails to load uses an architecture the engine does not support.
- Go back to the defaults. In Your defaults, select Whisper Large v3 Turbo (Q5) and Phi-3.5 Mini Instruct.
- Tell me. Settings, About, Copy Diagnostics gathers your Mac model, versions, model setup and latency stats, never your words, for a GitHub issue.
The short version
- Two models. Whisper Large v3 Turbo (Q5) hears you, and Phi-3.5 Mini Instruct cleans up.
- Both local. Bundled in the App Store build, and nothing is uploaded.
- Yours to swap. Pick from the libraries, or add any whisper.cpp or llama.cpp file.
On macOS 26 and later, the cleanup step can also use Apple Intelligence. Apple Intelligence and dictation on a Mac explains what that changes and what it does not.
The defaults are the ones I picked, and you can change either in two clicks. For AI dictation without the cloud in general, AI dictation on a Mac has more. If a model misbehaves, open an issue on GitHub. Every one gets read.
This app is completely free. You don't have to make an account. There are no subscriptions, no difficult downloads, and nothing is locked. Made for Apple silicon Macs. Source on GitHub.
Related
Questions
Does YapToText use Whisper?
Yes. It ships with Whisper Large v3 Turbo (Q5), the standard whisper.cpp file, running on your Mac. The Dictation Model Library offers ten other Whisper models, from Tiny to Large v3.
Which Whisper model should I use for dictation on a Mac?
Whisper Large v3 Turbo (Q5), the one YapToText marks Recommended and ships as the default. Smaller models are lighter on battery and memory, and Whisper Large v3 is the largest and slowest.
Is the cleanup done by a local LLM?
Yes. Phi-3.5 Mini Instruct, a llama.cpp model, runs on your Mac. You can swap it for Llama 3.2, Qwen 2.5 or Gemma 2 from the library, or add a GGUF file of your own.
Can I use my own Whisper or GGUF model?
Yes. Under Your own models, press Add Model and pick a whisper.cpp .bin or a llama.cpp .gguf file. The app tells which kind it is from the file itself.
How much disk space do the models take?
The two bundled models are 574 MB and 2.3 GB as the app shows them, most of the App Store download of about 2.9 GB. Library downloads show their size first and can be removed later.
Do the models need the internet?
No, not to run. The App Store build works offline from the first launch. The Homebrew build needs one speech model download, and model downloads are the app's only network use.