What Does a Dictation App Keep? A Privacy Checklist for Voice Typing (Audio, Transcripts, History and Keys)

Voice typing feels simple: press a key, talk, and text appears. Behind that one step, though, a dictation app may touch several kinds of data: your audio, the raw transcript, a cleaned-up version, a history log, and sometimes API keys for outside providers. If you pick a voice to text app for Windows, macOS or Linux for client work, it helps to know where each of those pieces goes and how long it stays there.

This is a practical privacy checklist for dictation apps. It is not legal advice, and every app's own privacy policy is the final word. It is a set of questions to ask before you trust a tool with your voice.

What "dictation data" actually means

A dictation app can create up to five kinds of data:

Data	What it is	Why it matters
Audio	The recording of your voice	It can contain anything you said, including names and numbers
Raw transcript	The words the speech engine heard	Shows exactly what you said, mistakes included
Formatted text	The cleaned-up version an AI model produced	May be sent to a second service for formatting
History	A log of past dictations	Convenient, but it piles up over time
Keys and vocabulary	API keys, custom words, corrections	Keys are credentials; vocabulary can reveal client names

Each of these can live on your device, on the app maker's servers, on a third-party speech provider's servers, or nowhere at all after use.

Question 1: Where does the audio go?

This is the big one. There are three common answers:

1. On your device only. Offline speech to text on Windows, macOS or Linux usually means a local model such as Whisper running on your own computer. The audio never needs an internet connection.
2. The app maker's servers. Many apps send audio to their own hosted speech models.
3. A third-party provider you choose. With "bring your own key" (BYOK), the app sends audio to a provider like OpenAI, Groq or Deepgram using your own API key, so that provider's data policy applies.

Built-in tools vary too. On Windows, voice typing is tied to the Online speech recognition setting under Privacy & security → Speech. On Apple silicon Macs, Apple says many languages are processed on the device. Check the current docs for your system, because these details change.

What to look for: a clear setting that shows which engine is active, and an offline option if you handle sensitive material.

Question 2: Is audio stored after transcription?

Sending audio for transcription is one thing. Keeping it is another. Look for a plain statement on whether recordings are saved, and for how long. "We process audio" is not the same as "we store audio".

What to look for: an explicit "no audio stored" statement, or a retention period you can live with.

Question 3: Where does the text history live, and can you turn it off?

History is useful. You can recover a long paragraph that landed in the wrong window. It is also a quiet archive of everything you dictated.

Ask:

• Is history stored locally, in the cloud, or both?
• Is there a limit on how many entries are kept?
• Can you clear it, or set it to keep nothing?

What to look for: a local history you can cap or switch off, and a cloud history with a stated time window and a clear button.

Question 4: Does AI formatting send text somewhere else?

Smart formatting, the feature that removes "um" and fixes punctuation, often runs on a separate language model. That means your transcript may travel to a second service after the speech engine is done.

What to look for: a choice of formatting model, the option to use a local model (for example through Ollama or LM Studio), and a fallback that pastes the raw transcript if the AI is unreachable, so you never lose what you said.

Question 5: How are API keys and vocabulary handled?

If you use BYOK, your API keys are credentials. If they sync across devices, they are stored on the app's servers. Custom vocabulary (client names, project codenames, medical or legal terms) can also be sensitive.

What to look for: masked keys, a test button, an easy way to remove keys, and the ability to keep sensitive saved text encrypted.

A quick checklist

Check	Good sign
Engine visible	The app shows which provider and model are active
Offline mode	A local model option, desktop at least
Audio retention	Stated plainly, ideally not stored
Local history	Capped, clearable, can be set to zero
Cloud history	Time-limited, text only, clearable
Formatting model	Selectable, local option available
Keys	Masked, removable, optional
Sensitive snippets	Encryption option for private saved text

How one app answers these questions

As an example, here is how InsertKit, a desktop app for voice typing, text snippets and clipboard history, handles the points above, based on its documentation:

• Engine choice: a Speech Engine panel with InsertKit Free, InsertKit Premium, BYOK providers (OpenAI, Groq Whisper API, OpenRouter, Deepgram, AssemblyAI, ElevenLabs Scribe) and Local Offline (Whisper). Status pills show the current provider, model and language.
• Offline mode: Local Offline Whisper runs fully on your computer, with models from Tiny (75 MB) up to Large v3 (3 GB). It is desktop only.
• Local history: Voice History is stored on the device only. The Voice Typing tab keeps up to 200 entries by default, the number is adjustable, and setting it to 0 turns saving off.
• Cloud history: Settings → Usage shows the last 30 days of voice text history, text only, no audio stored, with a Clear button.
• Formatting: you can choose the formatting model, add a custom OpenAI-compatible provider such as Ollama or LM Studio for local AI, and the raw transcript is pasted if formatting fails.
• Keys: the API Keys tab has Save, Test and Remove for each provider, with masked keys. Keys are optional, because InsertKit Free and Premium need none. Keys you add sync to your account.
• Encryption: private snippets can be end-to-end encrypted before they sync (a beta feature) with a passphrase that cannot be recovered. Team snippets stay unencrypted so teammates can use them.

It works in 100+ languages, including English and Bangla, on Windows, macOS and Linux. If those defaults fit how you work, you can download InsertKit for Windows, macOS or Linux.

FAQ

Is offline dictation always more private? It removes the network step for audio, which is a real difference. Your formatting settings still matter, because a cloud formatting model can receive the text.

Does BYOK make dictation private? Not by itself. It changes who receives your audio. The provider's own data policy then applies.

Should I turn off history? If you dictate sensitive material often, capping or disabling local history is a reasonable default. Keep it on if you rely on recovering lost text.

What is the fastest privacy win? Check which speech engine is active right now. Many people never look after the first setup.