Dictation glossary

Apple Neural Engine

Apple Neural Engine is a dedicated block of silicon on Apple chips that runs neural networks locally, alongside the CPU and GPU. Speech models compiled with CoreML can execute there so transcription happens on the device instead of in a remote API. It is hardware acceleration, not a speech model of its own. Windows and Linux machines typically run the same class of models on a CPU or GPU instead.

Get WhisperJot Try the browser demo

In more detail

What is Neural Engine?

The Neural Engine exists because matrix math for inference is wasteful on a general CPU if you do it all day. Apple exposes it through CoreML: a model is converted, compiled for the chip, and scheduled by the OS. For dictation that means a local recognizer can keep real-time factor under one without spinning up a data-center GPU. Battery and fan noise improve when the work stays on the accelerator. The first run may compile the model, which is why the initial dictation can feel slower than the tenth.

Not every on-device app uses the Neural Engine well. Some fall back to CPU if the model format is wrong, if the conversion failed, or if the process is sandboxed away from the accelerator. Intel Macs have no Neural Engine, which is why some Mac dictation tools require Apple silicon. On a Windows laptop the equivalent story is CPU cores or an NVIDIA GPU, not the ANE. When a spec sheet says 'Neural Engine,' read it as a Mac performance note, not as a feature Windows users are missing in name only.

For writers who speak

Why it matters for dictation

Local dictation is only pleasant if it keeps up with speech. The Neural Engine is how many Mac apps make on-device models feel immediate enough to replace typing. If your machine has no such accelerator, you still have on-device options — they just spend CPU — or you pick a cloud engine that moves the math off the laptop entirely.

In this product

How WhisperJot handles it

On Mac, WhisperJot's local engines run on Apple silicon (M1 and later) and can use the Neural Engine via CoreML; Intel Macs are not supported. Windows and Linux local engines run on the computer's CPU or GPU. Jot Cloud is the opt-in way to skip local acceleration and the model download on older or low-memory machines.

Related terms

Keep reading in the glossary.

On-device transcription

On-device transcription means speech recognition that runs on the computer or phone in front of you, using a model loaded in local memory.

Read On-device transcription →

Parakeet

Parakeet is a family of open speech recognition models from NVIDIA designed for fast automatic speech recognition.

Read Parakeet →

Transcription latency

Transcription latency is the time between speech and usable text — first partial, or the finalized insert after you stop.

Read Latency →

Offline dictation

Offline dictation means speech-to-text that keeps working with no internet connection because the recognition model runs on your computer.

Read Offline dictation →

Questions

Straight answers.

What is the Apple Neural Engine?

The Apple Neural Engine is specialized hardware on Apple chips for running neural networks on the device. Camera features, some system dictation, and third-party CoreML models can use it. For speech-to-text it is an accelerator, not the recognizer itself. The model still has to be compiled into a format the Neural Engine can execute. Other platforms use CPUs or GPUs for the same job.

Does on-device dictation require a Neural Engine?

No. On-device only means the model runs on that computer. A Windows or Linux PC can transcribe locally on a CPU. The Neural Engine is an Apple-silicon shortcut that often makes that local run faster and cooler. If a Mac app requires Apple silicon, it may depend on that accelerator or on instruction sets Intel Macs do not have.

What is CoreML in this context?

CoreML is Apple's framework for running machine-learning models on Mac and iOS, including on the Neural Engine, GPU, and CPU. Dictation apps that ship CoreML models can keep recognition on-device without writing GPU kernels by hand. CoreML is infrastructure. Whisper, Parakeet, or a voice activity model would be the payload it executes.

One hotkey, any focused app.

Private local transcription by default, with an optional opt-in cloud engine.

Get WhisperJot — $99/year