Speech and transcription models on Mac: Whisper and the apps that use it

Whisper models on a Mac are the speech-recognition files that transcription and dictation apps download so they can turn speech into text without sending audio anywhere. Each one is a single file or folder, from about 75 MB for the smallest to about 3 GB for each large version, and each stays on your Mac until you remove it.

Where they live depends on the app. MacWhisper, VoiceInk and Buzz keep them in folders of their own in your Library. Superwhisper manages them in its Models library. Command-line tools put them in hidden folders in your home folder, such as ~/.cache/whisper or the Hugging Face cache.

Removing a model you don't use is safe: the app downloads it again if you pick it later, and your transcripts don't depend on it. The best way is the app's own model list. The things to keep are your recordings and transcripts, which are often stored right next to the models and can grow larger than them.

Apple's own speech features, Dictation and Live Captions among them, are different. Their models belong to macOS, which downloads, shares and removes them by itself.

How big are Whisper models?

Whisper is OpenAI's open-source speech-recognition model, released in several sizes. A bigger size is more accurate and needs more memory and disk space. Most apps that transcribe offline on a Mac offer some of these sizes, often next to models from other makers.

The sizes, from tiny to large

OpenAI's README lists five sizes plus turbo, with the number of parameters each has. The file sizes below are what OpenAI's download server reports for its own files, and what whisper.cpp, the library many Mac apps build on, lists for its converted versions.

Size	Parameters	English-only version	OpenAI's file	whisper.cpp's file
tiny	39 million	tiny.en	76 MB	75 MiB
base	74 million	base.en	145 MB	142 MiB
small	244 million	small.en	484 MB	466 MiB
medium	769 million	medium.en	1.5 GB	1.5 GiB
large (v1, v2, v3)	1,550 million	None	3.1 GB each	2.9 GiB each
turbo (large-v3-turbo)	809 million	None	1.6 GB	1.5 GiB

The two columns differ a little because one counts in decimal megabytes and the other in binary mebibytes, and because the formats aren't identical. For deciding what to keep, they're the same size.

Why are there three large models?

OpenAI released large three times: large-v1, large-v2 and large-v3. Each is a separate file of about 3 GB. Asking OpenAI's package for "large" downloads large-v3 today, but a Mac that has used Whisper for a few years can hold all three.

They don't share anything. If you've kept large-v2 for one project and large-v3 for another, that's 6 GB for two versions of the same model.

What is turbo?

Turbo, also called large-v3-turbo, is an optimized version of large-v3. OpenAI describes it as faster, with minimal loss of accuracy, at about half the size: 809 million parameters against 1,550 million.

The catch is translation. OpenAI says turbo wasn't trained to translate speech into English, and returns the original language if you ask it to. If you translate, keep a multilingual model that isn't turbo; if you only transcribe, turbo can replace large.

What do the English-only models change?

The .en versions of tiny, base, small and medium only understand English. OpenAI says they tend to do better than the multilingual versions of the same size, especially tiny.en and base.en. They're the same size on disk, so choosing one saves nothing by itself, but it can let you pick a smaller size.

What are quantized models?

A quantized model stores each of its numbers in fewer bits, so the file is much smaller, at a small cost in accuracy. whisper.cpp offers quantized large models: large-v3 drops from 2.9 GiB to 1.1 GiB, and large-v3-turbo from 1.5 GiB to 547 MiB.

Apps label them in their own ways, such as "Quantized" in VoiceInk's list or a size in the name in WhisperKit's. If an app offers a quantized version of the model you use, it's the cheapest way to keep a large model.

• tiny0.1 GB
• base0.2 GB
• small0.5 GB
• large-v3-turbo, quantized0.6 GB
• large-v3, quantized1.2 GB
• medium1.5 GB
• large-v3-turbo1.6 GB
• large-v33.1 GB

Sizes from OpenAI's download server and whisper.cpp's model list (quantized models are whisper.cpp's q5_0 versions), converted to GB and rounded. Unquantized versions in other formats are close in size.

How much memory does a model need?

Disk space and memory are separate questions. whisper.cpp lists about 273 MB of memory for tiny and about 3.9 GB for large while a model runs. A model takes memory only while an app has it loaded, and disk space all the time, whether you use it or not.

That's why a large model you tried once and abandoned costs more than it seems. It used memory for one afternoon, and has used 3 GB of disk every day since.

One model, many formats

Whisper has been converted into several formats, each made for a different way of running it. The same model in two formats is two files of about the same size, and apps can't share each other's copies, so one Mac can hold the same model several times.

OpenAI's own files

OpenAI's Python package downloads PyTorch files named after the model, such as small.pt or large-v3-turbo.pt. It checks each file against a checksum, and downloads it again if the copy on your Mac doesn't match. Asking for "turbo" or "large-v3-turbo" gives the same file.

whisper.cpp's ggml files

whisper.cpp, a C and C++ version of Whisper that runs well on Apple silicon, uses its own format. Its files are named like ggml-base.en.bin, and it downloads them into a models folder. Many Mac apps are built on whisper.cpp, so a ggml- file in an app's folder is usually a Whisper model.

On Apple silicon, whisper.cpp can also use a Core ML version of the model's encoder, a folder named like ggml-base.en-encoder.mlmodelc, so part of the work runs on the Neural Engine. That adds a second item for the same model.

Core ML models and WhisperKit

WhisperKit, by Argmax, runs Whisper entirely as Core ML models, downloaded from a Hugging Face repository the project publishes. MacWhisper offers WhisperKit models next to its others, only on Macs with Apple silicon. Its documentation says the first load takes a while as the model is prepared for your Mac.

WhisperKit models come in folders rather than single files, some with their size in the name, such as openai_whisper-large-v3-v20240930_626MB. Where they're stored depends on the app that uses them.

MLX, faster-whisper and Transformers

Three Python libraries download Whisper from the Hugging Face Hub. mlx-whisper uses versions converted for Apple's MLX framework, published by the MLX community. faster-whisper uses versions in the CTranslate2 format. Hugging Face's own Transformers library uses OpenAI's published weights.

All three keep their downloads in the Hugging Face cache, where each model is a folder named like models--mlx-community--whisper-large-v3-turbo. The post on the Hugging Face cache explains how that cache is laid out and why it grows.

Format	File or folder looks like	Used by
PyTorch	large-v3.pt	OpenAI's whisper package
ggml	ggml-large-v3.bin	whisper.cpp and many Mac apps built on it
Core ML encoder	ggml-large-v3-encoder.mlmodelc	whisper.cpp on Apple silicon
Core ML, WhisperKit	openai_whisper-large-v3… folders	WhisperKit and apps built on it
MLX	models--mlx-community--whisper-…	mlx-whisper
CTranslate2	models--Systran--faster-whisper-…	faster-whisper

Distil-Whisper

Hugging Face also published Distil-Whisper, smaller models trained to behave like large ones. distil-large-v3 has about 756 million parameters, roughly half of large-v3, and its model card says it comes within 1% of large-v3's word error rate on long recordings.

It understands English only. If all your audio is English, it's a reasonable middle ground, and faster-whisper can load it by name.

Where Whisper models live on a Mac, app by app

Each app chooses its own folder, and most of them are hidden in your Library or in a folder whose name starts with a dot. This table covers the apps and tools people most often use. The sections after it give the details and the catch for each.

App or tool	Where its models go	Where to remove them
MacWhisper	~/Library/Application Support/MacWhisper/Models	Manage Models
Whisper Transcription (sandboxed MacWhisper)	The same folders inside ~/Library/Containers/com.goodsnooze.MacWhisper	Manage Models
Superwhisper	Not documented	Settings › Models library
VoiceInk	~/Library/Application Support/com.prakashjoshipax.VoiceInk/WhisperModels	AI Models › Local, on each model's card
Buzz	~/Library/Caches/Buzz	Help › Preferences › Models
OpenAI's whisper package	~/.cache/whisper	Finder: one file per model
mlx-whisper, faster-whisper, Transformers	~/.cache/huggingface/hub	Finder or Hugging Face's own tool

MacWhisper

MacWhisper's documentation says it keeps its data only in your user folder. Its main folder, ~/Library/Application Support/MacWhisper, holds three things: Models, the speech models you downloaded; Database, your transcripts and sessions; and External Media, media files kept with transcripts.

The sandboxed build, called Whisper Transcription, keeps the same three folders inside ~/Library/Containers/com.goodsnooze.MacWhisper. You download and switch models in the app's Manage Models window, where you can filter the list by engine.

Superwhisper

Superwhisper's model guide says you browse every model, download local ones and remove the ones you no longer use in Settings, under Models library. It doesn't say where downloaded models are stored on disk, so use that list rather than looking for files.

Its local models include Whisper Large v3 Turbo, 1.6 GB, alongside models from other makers, such as Parakeet Multilanguage at 494 MB. Superwhisper also keeps a folder of its own, covered below under recordings.

VoiceInk

VoiceInk lists its local models under AI Models › Local, including Whisper tiny, base, large-v2, large-v3 and large-v3 turbo, with a quantized turbo. Its documentation says each model card's menu deletes the downloaded files or shows them in Finder.

VoiceInk's source code puts Whisper models in a WhisperModels folder inside ~/Library/Application Support/com.prakashjoshipax.VoiceInk. The same folder holds its recordings and its transcript history, so don't move the whole folder to free space.

Buzz

Buzz, an open-source transcription app, keeps its models in ~/Library/Caches/Buzz, its FAQ says. It offers several engines, including OpenAI's Whisper, whisper.cpp, faster-whisper and models from Hugging Face. Its FAQ recommends whisper.cpp on a Mac.

To see the files, the FAQ points to Help › Preferences › Models, where Show file location opens the folder after you've downloaded a model. Because the models sit in a Caches folder, anything that clears your caches removes them too; Buzz downloads them again when you transcribe.

Aiko

Aiko, a transcription app for Mac and iPhone, uses the Whisper large-v2 model on a Mac, its website says. It offers no list of models to download or remove, and the App Store lists the app at about 1.8 GB. Its makers say you can't remove single languages to save space.

Command-line tools

If you've used Whisper from Terminal or from a script, the models are wherever the tool put them:

• OpenAI's whisper package keeps one file per model in ~/.cache/whisper. If you've set the XDG_CACHE_HOME variable, the whisper folder is inside that folder instead.
• whisper.cpp saves models in the models folder of wherever you downloaded whisper.cpp, named like ggml-base.en.bin.
• mlx-whisper, faster-whisper and Transformers use the Hugging Face cache, ~/.cache/huggingface/hub.

The OpenAI package also depends on PyTorch, which is large in its own right. The post on how much storage local AI models need covers what Python environments and model runtimes add on top of the models themselves.

How to find Whisper models you've forgotten about

Storage settings doesn't name speech models. Choose Apple menu › System Settings, click General, then Storage: models in hidden folders and in your Library usually count toward System Data, with nothing to say which app they belong to. Finder finds them faster.

How do you open hidden folders like ~/.cache?

In Finder, choose Go › Go to Folder, type ~/.cache, and press Return. If Finder can't find it, nothing has created that folder. To see hidden folders in your home folder as you browse, press Command-Shift-Period in a Finder window; press it again to hide them.

The same Go to Folder step opens ~/Library/Application Support and ~/Library/Caches, where the apps above keep their folders. Both sit in your Library, and the post on the Mac Library folder explains what else is inside it.

How do you search for model files?

Most Whisper models have recognizable names. In a Finder window, press Command-F, click This Mac, and search for ggml or whisper. Files named ggml- and ending in .bin, folders ending in -encoder.mlmodelc, and files like medium.pt are almost always speech models.

To be sure, check ~/.cache and the app folders above by hand as well.

How do you see how much space each one takes?

Select a file or folder and choose File › Get Info. Or open the folder, choose View › as List, then View › Show View Options, and select Calculate all sizes so folders show their size too. Click the Size column to put the largest first.

Which of your apps downloaded them?

The folder name usually says. A model inside an app's own folder belongs to that app, and only that app uses it. Copies of the same model in two apps' folders aren't shared: each app downloaded its own, which is a common reason the same model takes space twice.

Is it safe to delete Whisper models?

Yes, for models you don't use. A downloaded model is a copy of a file its publisher still offers, so removing it loses nothing you made. The catalog entry for OpenAI's package, OpenAI Whisper models, puts the cost plainly: the next transcription with a removed model downloads it again first.

What do you lose when you remove one?

Only time and bandwidth. Picking a removed model again means downloading it again, about 3 GB for a large one. Some engines then prepare the model for your Mac on first use, as MacWhisper's documentation says WhisperKit models do, so the first transcription is slower.

If you remove the model an app is set to use, the app needs it back, or another model, before it can transcribe. Choose your replacement first, then remove the old one.

Do your transcripts depend on the model?

No. A transcript is text, saved in the app's database or as a file, and nothing in it needs the model that made it. The audio is different: if you want to transcribe a recording again with a better model later, you need the recording, not the old model.

Which files shouldn't you remove?

• A model you imported. VoiceInk can import a whisper.cpp file, and other apps can too. If you made or converted it yourself, it may not download again, so keep the original somewhere first.
• An app's whole folder. MacWhisper's and VoiceInk's main folders hold your transcripts and recordings beside the models. Remove a model, never the folder around it.
• Anything inside an app's database. Delete transcripts in the app, so it keeps its records in step.

Does deleting a model free space right away?

Not until the Trash is emptied, if you removed it in Finder. If you removed it in an app's own list, the app decides; check your free space afterwards, and empty the Trash if the model went there.

What you remove	Safe?	What it costs
A model you don't use, in the app's list	Yes	A download if you want it again
The model the app is set to use	Yes, after choosing another	The app can't transcribe until you choose
A model you imported or converted yourself	Only with a copy elsewhere	It may not download again
A model file in ~/.cache/whisper	Yes	The package downloads it again when asked
An app's whole Application Support folder	No	Transcripts, recordings and settings

How to delete Whisper models in each app

Use each app's own list wherever it has one. The app then knows the model is gone, and you don't have to work out which files, such as Core ML folders, belong to it.

In MacWhisper

1. Open MacWhisper and open the Manage Models window.
2. Find the models you've downloaded and no longer use.
3. Remove them there, if your version offers it.

MacWhisper's documentation doesn't describe a remove step, but says all its data can be inspected or removed by hand. If your version has no button, quit MacWhisper, open its Models folder with Go › Go to Folder, and move only the model you don't use to the Trash.

In Superwhisper

1. Open Superwhisper's Settings.
2. Click Models library.
3. Remove the local models you no longer use.

Check the modes you use afterwards. Superwhisper's guide says choosing a local model in a mode downloads it on first use, so give any mode that used a removed model the one you kept.

In VoiceInk

1. Open VoiceInk and click AI Models, then Local.
2. On the card of a model you don't use, open the menu.
3. Choose to delete its downloaded files.

The same menu can show the files in Finder, which is a quick way to check their size first.

In Buzz

1. Open Help › Preferences and click Models.
2. Select a downloaded model.
3. Delete its files there, as Buzz's FAQ describes for a damaged download.

For OpenAI's whisper package

There's no app for it, so this one is done in Finder.

1. Choose Go › Go to Folder, type ~/.cache/whisper, and press Return.
2. Choose View › as List to see each file's size.
3. Move the files of models you don't use, such as large-v2.pt, to the Trash.
4. Empty the Trash when you're sure.

Removing the whole whisper folder is fine too; the package creates it again when it next downloads a model.

For the Hugging Face cache

Moving one model's folder, such as models--mlx-community--whisper-large-v3-turbo, to the Trash removes that model and nothing else. The library downloads it again when a script asks for it.

Newer versions of Hugging Face's library may keep the files in a shared folder, so the model's own folder frees little; the Hugging Face cache post shows how to tell. Hugging Face's command-line tool can also remove single models and old revisions; it has no app.

Recordings and transcripts: the other half of the space

Models are a fixed cost: a few gigabytes, once. Recordings grow every day you dictate or transcribe, and several apps keep the audio of every session unless you tell them not to. On a Mac used for meetings or interviews, they can outgrow the models.

How much space does recorded audio take?

It depends on the format, which each app chooses. As a guide, Whisper's own code converts audio to 16,000 samples a second, in one channel. At that rate, an hour of uncompressed 16-bit audio is about 115 MB, so a few hours a week add up to gigabytes in a year.

Apps that keep the original file, such as a video you imported, keep it at its full size. MacWhisper's External Media folder holds media files it keeps with transcripts.

VoiceInk

VoiceInk keeps recordings in a Recordings folder next to its models. Its History Settings, opened from the button beside the search field in History, offer two kinds of cleanup:

• Auto-delete Audio Files removes only the audio after 1, 3, 7, 14 or 30 days, and keeps the transcript text.
• Auto-delete Transcript History removes transcripts, with their audio, after a period from immediately to 7 days.

Run Cleanup Now deletes items older than the period you chose right away. Deleting a transcript in History also deletes its audio.

Superwhisper

Superwhisper keeps its configuration and history in a folder of its own, ~/superwhisper by default. Its documentation lists a recordings folder there with your dictation history, beside modes, settings and models. Installs from before the default changed may use ~/Documents/superwhisper instead.

That older place matters if Desktop & Documents Folders is on in iCloud Drive. Superwhisper's documentation says iCloud syncing can move the folder, and then new modes fail to save. Its Advanced settings show the folder in use, with a Change folder button.

MacWhisper and Aiko

MacWhisper keeps transcripts in its Database folder and media in External Media, both beside its models. Manage them in the app, never by deleting files from the database.

Aiko's website says recordings aren't deleted automatically after transcription. Choose File › Show Recent Recordings to open its folder, and look in its settings for the option to delete recordings older than 7 days.

What should you keep?

Keep the recordings you might need again: interviews, lectures, anything you'd transcribe again with a better model one day. Dictation snippets are usually worth nothing once the text is pasted. Turning on an app's audio cleanup for dictation, and saving important recordings elsewhere yourself, keeps both under control.

Apple's own speech models, managed by macOS

Not every speech feature on a Mac downloads a model you can see. macOS has its own speech recognition, and apps can use it instead of shipping Whisper. Those models belong to macOS, which decides when to download and remove them.

Dictation

macOS Dictation turns speech into text in any app. To turn it on in macOS 27 Golden Gate, choose Apple menu › System Settings, click Keyboard, go to Dictation and turn it on, then click Enable. The steps are the same in macOS 26 Tahoe.

Once it's on, the text below Dictation in Keyboard settings says whether your voice input for general text dictation is processed on your Mac and not sent to Siri's servers. There's no list of dictation models to manage; macOS handles them.

Live Captions

Live Captions, in System Settings › Accessibility › Live Captions, captions any audio on your Mac. Apple says it's available only on Macs with Apple silicon, processes audio on the device, and downloads language data the first time you turn it on. That data is macOS's to manage, not a file you remove.

Apps that use Apple's speech framework

Since macOS 26, apps can transcribe with Apple's own speech models through the Speech framework. Apple's developer documentation says these models are downloaded from Apple's servers and managed by the system. Once one app downloads a language's model, macOS keeps it updated and shares it with other apps.

Each app can reserve only a limited number of languages, to limit storage and network use. When an app releases a language, macOS removes the model later; it can also stop holding models for an app that hasn't used them in a while. VoiceInk's model list, for example, includes an Apple Speech option that needs macOS 26 or later.

Do Apple's speech models take space you can free?

Not directly. There's no supported way to delete them, and you don't need one: macOS shares a single copy among apps and removes what no app has reserved. If space is tight, a dictation app that uses Apple's models adds nothing of its own, while one that uses Whisper adds up to about 3 GB for each model you choose.

Feature	Whose model	Can you remove it?
Dictation	macOS	No; macOS manages it
Live Captions	macOS, downloaded on first use	No; macOS manages it
Apps using Apple's Speech framework	macOS, shared between apps	No; macOS removes what no app reserves
Apps using Whisper or other models	The app	Yes, in the app's model list

Which Whisper model do you need on a Mac?

The smallest model that gives you text you don't have to fix. Keeping one model you trust, rather than four you tried, is the biggest saving most people can make. The table matches common uses to a sensible choice; try it on your own audio before you remove the rest.

What you do	A sensible model to keep	Why
Short English dictation	base.en or small.en	English-only models do well at small sizes
Dictation in several languages	small or turbo	Multilingual, without large's size
Meetings and interviews, English	turbo or distil-large-v3	Near large-v3, at about half the size
Transcribing many languages	turbo, or large-v3 if accuracy matters most	Turbo transcribes; large handles harder audio
Translating speech into English	large-v3 or medium	Turbo isn't trained to translate
A Mac with little free space	A quantized turbo	The smallest way to keep a large model

A worked example

Say you've tried three apps over two years. MacWhisper holds large-v3 and medium, VoiceInk holds large-v3 turbo and base, and an old script left large-v2 and small in ~/.cache/whisper. That's roughly 3.1, 1.5, 1.6, 0.15, 3.1 and 0.48 GB: about 10 GB of speech models.

You now use only VoiceInk with turbo. Removing the rest in MacWhisper's Manage Models and the two files in Finder brings speech models down to about 1.6 GB. If you later want MacWhisper again, it downloads one model, not all of them.

Does a bigger model mean better results?

Usually, up to a point. OpenAI's relative speeds were measured on a data-center graphics card, not a Mac, so they show the order rather than the time on your machine. Accuracy also varies widely by language, OpenAI says, so a size that's fine for English may not be for another language. Test on your own recordings before you settle.

What about models that aren't Whisper?

Apps now offer other speech models too, such as NVIDIA's Parakeet, Cohere Transcribe and SenseVoice. In the apps above, they're listed and removed alongside the Whisper models. The post on AI models hidden inside Mac apps compares them with the other models apps download for themselves.

Keeping speech models from piling up

Every model you try stays until you remove it. A few habits keep them to the one or two you use.

1. Pick a model for each app and remove the others in its list once you're happy with it.
2. When you stop using a transcription app, remove its models in the app before you delete it.
3. Turn on audio cleanup for dictation in apps that offer it.
4. Keep important recordings in a folder you back up, not only in an app's database.
5. Every few months, check ~/.cache and the app folders above for models you no longer recognize.

Mistakes to avoid

Mistake	Why it hurts	Instead
Moving an app's whole Application Support folder to the Trash	It holds your transcripts and recordings	Remove models in the app
Deleting the model an app uses	The app can't transcribe until you choose again	Choose a new model first
Keeping every large version	Each is about 3 GB, and they share nothing	Keep the newest one you need
Clearing caches and expecting Buzz to keep its models	They're in its Caches folder	Expect a download next time
Uninstalling an app without its models	Its folder stays in your Library	Remove models first, or remove its folder after

What happens to models when you delete the app?

They stay. Dragging an app to the Trash removes the app, not its folders in your Library, so MacWhisper's or VoiceInk's models and recordings remain until you remove them. Check the app's folder before you remove it: recordings you want to keep go somewhere safe first.

How the app does it

Wombit is a native Mac app that shows what's using your storage and explains every item in plain language. It scans the whole disk or any folder, reading only names, sizes and dates, and file names and paths never leave your Mac.

On the map, speech models show up with their real size wherever they are. Select one to see what it is, who made it and what removing it costs, from the catalog entries below.

• ~/.cache/whisper

  OpenAI Whisper models

  Speech recognition models downloaded by OpenAI's open-source Whisper package, one file per model size.

  Usually safe · Moves it to the Trash

• ~/.cache/huggingface/hub

  Hugging Face cache

  Models and datasets downloaded from Hugging Face, shared by every library and tool that uses them.

  Review first · Moves it to the Trash

• ~/Library/Application Support/*

  Application Support folders

  Data that apps create and manage for you, such as libraries, databases, downloads and settings, in a folder for each app.

  Review first · Explained only: the app shows you how

OpenAI's Whisper models are Usually safe: they download again when you transcribe, so nothing is lost, but they never start checked. The Hugging Face cache is Review first, because it holds every model your Python tools downloaded, not only speech models.

An app's own folder in Application Support, such as MacWhisper's or VoiceInk's, is Review first too: the catalog says to manage that data in the app, because an app's folder can hold things you made, such as transcripts and recordings.

Folders inside it that no rule describes get a Best guess, written on your Mac by Apple Intelligence and labelled as a guess.

Seeing what your models take is free. Moving them to the Trash, with a review sheet and a log, is PRO: one payment of $4.99 with lifetime access. Every Pro tool's review is free to open, so you see what would go and why before you buy.