macOS Dictation is Apple's built-in speech-to-text tool. The audio you speak is sent to Apple's servers for transcription — it does not stay on your machine. Apple encrypts the audio in transit and states that the audio is not used to train models, but uploading audio to a server, even an encrypted one, is fundamentally different from processing it locally.
What happens to your audio
When you dictate into macOS Dictation, the audio is transmitted to Apple's servers where the transcription engine processes it and sends back the text. Apple's privacy documentation describes the audio as encrypted during transmission and states it is not used for model training or sold to third parties.
The encryption protects the audio while in transit, but the content itself still leaves your device. For a use where audio must remain completely on the machine — medical, legal, financial, or restricted environments — this is the wrong tool. A server-based approach is fundamentally different from on-device processing, even with strong encryption.
The offline requirement
macOS Dictation requires an internet connection for every transcription. The tool does not work on a plane, in SCIFs, on locked-down corporate laptops with network restrictions, or in any environment where the device cannot reach Apple's servers. This is not a minor limitation — it is architectural. Every time you dictate, the audio must travel to Apple's infrastructure and return as text.
If offline capability is a requirement, on-device speech recognition is the relevant category. On-device tools run the transcription engine locally, so nothing leaves your device.
Understanding the retention policy
Apple's privacy documentation states that audio is not retained for model training. But what happens to the audio after transcription is not detailed. Some services delete audio immediately after creating the transcript; others retain it temporarily for quality purposes; still others keep it indefinitely. If audio retention after the transcription is complete matters to your use, check whether Apple documents that approach.
The gap between "not used for training" and "deleted immediately" is significant. Many services use audio for purposes other than training — testing model improvements, studying error patterns, or improving transcription across languages. Knowing what Apple does with audio after the transcript is created would matter if your conversations are sensitive.
Local processing as an alternative
If the prospect of audio reaching Apple's servers is unacceptable, on-device tools exist precisely to avoid it. The trade is straightforward: local processing means no upload, no network dependency, and no question about retention — the audio never leaves your machine in the first place. But it also means the transcription engine is limited to what fits on your device, and updates require you to download new model weights.
Where to go if server-based is not acceptable
The trade between encryption and server processing is legitimate. Apple's implementation is straightforward about what it does. If your use case requires that no audio leaves your machine under any circumstances, offline dictation apps for Mac covers the tools that handle it differently — processing everything locally and uploading nothing.
Many commercial dictation apps — including Wispr Flow — also send audio to servers for transcription, but with different retention and training policies. The comparison between Apple Dictation and third-party tools covers what a paid tool adds beyond Apple's free option.