Whipscribe › Answers
Answer

What is the best on-device speech-to-text for a Mac with Apple Silicon?

For a ready-made app, MacWhisper (local models, Pro is a one-time €64) or Aiko ($24 one-time, Whisper large v2 on macOS) are the usual picks; Buzz is the free, open-source one and needs Apple silicon. To build it into your own software, use WhisperKit for Swift apps (Core ML, macOS 14 or later) or whisper.cpp for C/C++ and the command line (Metal and the Neural Engine). All of these run on the Mac itself. WhipScribe is hosted, not on-device.

Price
Free preview of the opening minutes; packs $4 = 500 min · $12 = 2,000 min · $24 = 5,000 min; credits never expire
Per file
Up to 10 hours or 5 GB per file
Languages
99+, detected automatically
Turnaround
About 2 minutes per hour of audio; most files finish in under a minute
Checked
2026-10-25

Apps, if you just want to drop a file in

MacWhisper transcribes with local models by default ("Process sensitive content locally without data ever leaving your Mac") and sells Pro as a one-time licence of €64 with lifetime updates.

Aiko transcribes on the device and costs $24 once, with a 14-day trial through TestFlight. Its site says it uses the Whisper large v2 model on macOS.

Buzz is free and open source (MIT). Its README says the Mac version now requires Apple silicon, so it is not for an Intel Mac.

Libraries, if you are writing code

WhisperKit, from Argmax, is Swift and Core ML: "On-device Speech AI for Apple Silicon", MIT licensed, with macOS 14.0 or later and Xcode 16 as package prerequisites. It is the natural fit inside a Mac or iOS app.

whisper.cpp treats Apple Silicon as a "first-class citizen": inference runs on the GPU through Metal, and the encoder can run on the Apple Neural Engine through Core ML. It is MIT licensed and also builds for Linux, Windows and Raspberry Pi.

Memory is the practical limit. whisper.cpp lists the large model at 2.9 GiB on disk and about 3.9 GB of memory, medium at 1.5 GiB and about 2.1 GB, small at 466 MiB and about 852 MB.

When on-device is not the right answer

On an Intel Mac, an 8 GB machine, or a recording several hours long, a local model can be slow. A hosted service returns about two minutes per hour of audio. WhipScribe is one: one file up to 10 hours or 5 GB, 99+ languages, a free preview of the opening minutes and packs from $4 for 500 minutes. The audio is uploaded to WhipScribe's own servers, never to a third-party AI service, and is never used for training, but it does leave the Mac.

Sources

Related questions

Is MacWhisper or Aiko better?

Both run Whisper on the Mac. Aiko is $24 once with few settings; MacWhisper Pro is €64 once with more features, including optional cloud services with your own API keys. Try Aiko's trial first if you only need plain transcripts.

Do these apps use the Neural Engine?

whisper.cpp can run the encoder on the Apple Neural Engine through Core ML, and WhisperKit is built on Core ML. For the apps, check each app's own settings; their sites do not all say which chip does the work.

Which free option runs on an Intel Mac?

whisper.cpp still lists Intel Macs. Buzz now requires Apple silicon on the Mac. Or use a hosted service in the browser.

Is WhipScribe an on-device app?

No. WhipScribe is hosted: it runs in the browser and transcribes on its own servers. Use it when the Mac is slow or the file is long, not when audio must stay on the machine.

Keep reading

WhisperKit whisper.cpp Is MacWhisper worth it? Transcription apps for Mac Can you transcribe this audio? Summarize this recording. I have a 3-hour interview, can you analyze it?
Try it on your own recording

Every recording starts with a free preview of its opening minutes; packs from $4 for 500 minutes transcribe the rest. A single file may be up to 10 hours and 5 GB.

Transcribe a file →

Looking for a service?

For the recording you have in front of you: see how our transcription services work.