On-device speech recognition

On-device speech recognition converts speech to text with a model stored on your own hardware. The audio is processed on the device, so your words do not have to travel to a company's server to be transcribed. Apple says that, on a Mac, voice inputs and transcripts for general text Dictation are processed on your device and not sent to Siri servers ¹, and that you can check this in Keyboard settings.

The recognition model has to live on the device

A cloud service keeps a large recognition model on its own computers. An on-device recognizer needs a model small enough to store and run on a phone or laptop. In 2019 Google announced an end-to-end, all-neural, on-device speech recognizer to power speech input in Gboard ². Google also said the new recognizer is always available, even when you are offline ³. That was Google's 2019 description of its first on-device model, so later models may differ.

An example: dictating with no network

Suppose you draft a clinic note on a laptop in a train tunnel with no signal. With on-device recognition, the words still appear as you speak, because the model runs locally and nothing has to be sent out. A cloud dictation service in the same place either stops or queues the audio until the connection returns. The example shows the design, not a test of any product.

Why it matters for dictation

Dictation sends the most personal kind of input you produce: your own voice, often with names, addresses and medical or legal details. Where recognition happens decides whether that audio leaves your device. Microsoft describes its embedded speech feature as designed for on-device speech to text where cloud connectivity is intermittent or unavailable ⁴. Offline operation is a practical benefit too, because the text appears without waiting on a network round trip.

On-device recognition is not automatically the most accurate option. A model that must fit on a phone can be smaller than a server model, so the question of accuracy is worth checking for your language. The question on whether on-device speech recognition is as accurate as cloud covers that trade-off.

Check the setting, not only the label

Vendors use "on-device," "offline" and "local" loosely, and a product can use the on-device path for some languages and the server for others. Apple's help page notes that whether a language is processed on the device depends on the language and the Mac, so check the setting for your own language. Before you rely on a tool with sensitive material, try it with the network off and read the privacy policy for where audio is sent.

Yap, which we make, runs its dictation on the device on Mac and Windows. Compare that with any other tool by reading its settings and its policy.

Related terms and guides

On-device recognition is one kind of automatic speech recognition. The opposite design is cloud dictation, where audio goes to a remote service. A voice profile stores settings and custom words for one user and is separate from where recognition runs.

To go deeper, read is voice to text safe? How to dictate privately, which tests where your words go on each platform. For the rules that apply to voice data, see is your voice biometric data?.