Which dictation apps can use your recordings to train AI?

Two apps in our data say your recordings can be used to improve the maker's models, and both give you a choice. Microsoft says SwiftKey may send anonymized voice clips to Microsoft to improve recognition, and that you can opt out in SwiftKey settings ¹. Google says you can choose to donate snippets of your speech to improve its speech models ².

The data comes from makers' own pages: 1676 data points ³ from 583 sources ⁴. A 1 means the page says the app has the feature. The sparse list of training rows says less than it seems to, as the sections below explain.

Training rows are rare, and both are opt-in or opt-out

Showing 8 of 107 rows. See every row in the dataset.

The SwiftKey sentence says the clips improve recognition, not that they train a named model. Microsoft's article does not say how long clips are kept. Gboard's donation is opt-in. Google's page also says federated learning is on by default, but that it does not send the text you speak or type to Google.

The difference between opt-in and opt-out matters. Under opt-in, nothing is shared until you act. Under opt-out, sharing is the starting state and you have to find the setting.

Most dictation apps say nothing about training

Most apps in the dataset have no training row at all. A page that does not mention training gives no row, so silence is not the same as "no training". The rows shown here are all yes, so they do not list apps that deny training. To learn what a specific app does, read its privacy policy and data controls yourself.

Audio kept on servers is the nearer risk

Training needs stored data, so where audio lives matters. Three makers say they keep audio or data on their servers.

AudioPen's privacy policy says audio from successfully processed recordings is deleted from its servers within 48 hours ¹¹. Its FAQ says voice typing audio is not stored at all, so this covers voice notes.

Paraspeech says it stores feedback audio that users choose to send ¹². This happens only when you select Include recording, which is off by default, and dictation audio is not stored this way.

Otter's security page says it uses AWS S3 storage for data, with server-side encryption on the buckets ¹³. The sentence says "data", not audio, and we read it as covering meeting recordings.

Showing 8 of 213 rows. See every row in the dataset.

Text often goes to a third-party AI even when audio does not

Many apps never send your recording out, but they do send the transcript to an AI provider for rewriting. This is a separate row because the risk differs: the provider receives words, not your voice.

WhisperTyping's privacy policy says the text of its AI features is sent to OpenAI (1 for yes) ²¹. That applies to the AI modes only. BetterDictation says its Pro add-on uses OpenAI to clean transcripts ²².

Others let you choose the provider. Voicetypr says transcribed text and formatting instructions go to the provider you configure, and that the original recording is not sent in that request (1 for yes) ²³. Glimpse says that with an optional LLM feature, the text and prompt go to the provider you set up ²⁴. DictaFlow names OpenAI, Deepgram and Groq as providers that may be used ²⁵.

Showing 8 of 177 rows. See every row in the dataset.

Whether a provider keeps or trains on that text depends on the provider and your account with it. The maker's page usually does not say.

What the numbers do and do not measure

These rows are claims on makers' pages, read in October 2026. They are not audits and we did not watch network traffic. A policy can change, and a "may include" list does not say which provider handles a given request. Reviews and third-party comparisons are left out. Some rows apply to one plan, one mode or one platform, so read the detail column.

What this means for choosing a tool

1. Check opt-in or opt-out first. For keyboards from large platforms, look in the settings for voice data sharing.
2. Prefer speech on the device if recordings worry you. If audio never leaves your machine, there is nothing to store or train on. Our page on offline dictation apps shows who says so.
3. Ask where the text goes. AI cleanup that uses your own key sends text to the provider you pick.
4. Read the policy for the plan you use. Free and paid tiers can differ.

Yap, which we make, is on-device dictation for Mac and Windows, and its own pages describe its data handling. For how voice data is handled more broadly, read voice privacy. For the wider picture, see the overview of dictation software, the guide to the best dictation software and the market overview.

Every row is in the full dataset.