whisper.cpp vs MLX Whisper vs faster-whisper: which Whisper to run on a Mac
By Ryleigh Newman · Published
On an Apple silicon Mac, the two Whisper implementations built for the hardware are whisper.cpp and MLX Whisper. whisper.cpp is plain C and C++ with no dependencies, and its README says inference runs fully on the GPU through Metal. MLX Whisper runs on MLX, Apple's machine learning framework for Apple silicon, and installs with pip. faster-whisper is the third name people compare: it is fast on NVIDIA GPUs and on CPUs, and its README lists no Apple GPU path.
Every fact here is from the three projects' own READMEs on GitHub, read on October 11, 2026. Speed figures are theirs, not mine. I built YapToText, a free Mac dictation app that uses whisper.cpp, so I have a stake in one of the three.
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: 8 sections
Side by side
From each project's README on GitHub, read October 11, 2026.
| whisper.cpp | MLX Whisper | faster-whisper | |
|---|---|---|---|
| Written in | C and C++, no dependencies | Python, on Apple's MLX framework | Python, on the CTranslate2 engine |
| Install | Clone and build with CMake | pip install mlx-whisper | pip install faster-whisper |
| Apple silicon GPU | Yes, through Metal. Core ML for the encoder is optional | Yes. MLX is built for Apple silicon | Not listed. Its GPU section covers NVIDIA libraries |
| Intel Macs | Yes. The README lists Mac OS on Intel and Arm | No. MLX is for Apple silicon | CPU |
| Needs FFmpeg | Only to convert input. Its command line tool reads 16-bit WAV | Yes | No. It decodes audio with PyAV |
| Model format | ggml files, with integer quantization | MLX weights, with a 4-bit option | CTranslate2 models, with 8-bit quantization |
| Word timestamps | Experimental | Yes | Yes |
| License | MIT | MIT | MIT |
whisper.cpp
whisper.cpp describes Apple silicon as a first-class citizen, optimized through ARM NEON, the Accelerate framework, Metal and Core ML. It runs on Mac, iOS, Android, Linux, Windows, WebAssembly and a Raspberry Pi, which is why so many apps embed it.
git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
sh ./models/download-ggml-model.sh base.en
cmake -B build
cmake --build build -j --config Release
./build/bin/whisper-cli -f samples/jfk.wav- Core ML. On Apple silicon the encoder can run on the Apple Neural Engine. The README says that can be more than three times faster than CPU-only execution, and that generating the Core ML model needs Python tools and Xcode.
- Quantization. Models can be quantized, for example with the Q5_0 method, to take less memory and disk.
- Input. The command line tool takes 16-bit WAV, so convert other files with ffmpeg first.
- Memory. The README lists about 273 MB for tiny, 852 MB for small, 2.1 GB for medium and 3.9 GB for large.
The local Whisper guide walks through this setup in full.
MLX Whisper
MLX is Apple's array framework for machine learning on Apple silicon, and mlx-whisper is the Whisper example built on it. It is the shortest route from nothing to a transcript if you already have Python.
brew install ffmpeg
pip install mlx-whisper
mlx_whisper audio_file.mp3- Output. That command writes audio_file.txt. A flag sets other formats.
- Models. The default is the tiny model. Pre-converted models are in the MLX Community on Hugging Face and download automatically when you name one.
- In Python. One call, mlx_whisper.transcribe, returns the text, with word-level timestamps on request.
- Quantized models. Its conversion script can make a 4-bit model.
faster-whisper
faster-whisper reimplements Whisper on CTranslate2. Its README says it is up to four times faster than the original Python implementation for the same accuracy, with less memory, and its benchmark tables were run on an NVIDIA RTX 3070 Ti and an Intel Core i7. Those are its numbers, on that hardware.
For a Mac, the relevant line is in its requirements: GPU execution needs NVIDIA's cuBLAS and cuDNN libraries. A Mac has no NVIDIA GPU, so on a Mac it runs on the CPU. It is still a good choice when the same Python code has to run on a Linux server with a GPU as well.
Which to pick on a Mac
My reading of the READMEs above.
| If you | Use |
|---|---|
| Want a transcript from the Terminal in two commands | MLX Whisper |
| Are embedding Whisper in an app, or need Intel Macs too | whisper.cpp |
| Need the smallest files and memory use | whisper.cpp with a quantized model |
| Write Python that also runs on NVIDIA servers | faster-whisper |
| Do not want a Terminal at all | An app that bundles one of them |
None of the three READMEs gives a like-for-like speed comparison on Apple silicon, and I am not going to invent one. Try your own audio on your own Mac.
What YapToText uses
YapToText runs whisper.cpp. Its App Store download includes Whisper Large v3 Turbo as the stock 5-bit quantized whisper.cpp file, 574 MB, so dictation and file transcription work with no build step, no Python and no model download. It needs an Apple silicon Mac with macOS 14 or later. If you want to try the same model by hand first, the Whisper Large v3 Turbo page covers the file, and the Whisper Mac apps page lists other apps that wrap Whisper. Parakeet vs Whisper compares Whisper with the other model family people run locally.
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.
The short version
- whisper.cpp. C and C++, Metal and optional Core ML, Intel and Apple silicon, the one most apps embed.
- MLX Whisper. Python on Apple's MLX, Apple silicon only, two commands to a transcript.
- faster-whisper. Python on CTranslate2. Its GPU path is NVIDIA, so CPU on a Mac.
- All three. MIT licensed.
If one of these projects changes, open an issue on GitHub and I will update this page. 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
Sources
- https://github.com/ggml-org/whisper.cpp (checked 2026-10-11)
- https://github.com/ml-explore/mlx-examples/tree/main/whisper (checked 2026-10-11)
- https://github.com/ml-explore/mlx (checked 2026-10-11)
- https://github.com/SYSTRAN/faster-whisper (checked 2026-10-11)
- https://github.com/ryleighnewman/YapToText (checked 2026-10-11)
- https://apps.apple.com/us/app/yaptotext/id6786382289 (checked 2026-10-11)
Questions
Is MLX Whisper faster than whisper.cpp?
Neither project's README gives a like-for-like comparison on Apple silicon. Both use the Mac's GPU, so test your own audio on your own Mac.
Does faster-whisper use the GPU on a Mac?
Its README lists NVIDIA libraries for GPU execution. A Mac has no NVIDIA GPU, so it runs on the CPU there.
Does whisper.cpp use Metal?
Yes. Its README says that on Apple silicon the inference runs fully on the GPU through Metal, with Core ML as an option for the encoder.
How do I install MLX Whisper?
Install ffmpeg, then run pip install mlx-whisper. The command mlx_whisper followed by a file name writes a text transcript.
Which Whisper implementation works on Intel Macs?
whisper.cpp lists Mac OS on Intel and Arm. MLX is for Apple silicon only.