Short version: every voice typing app is features its team shipped plus features it chose not to ship. The magical demos usually need more access than the marketing page shows. Use the matrix, then the sections below. This is not legal advice; verify each vendor's current policy before you install.
Comparison matrix (verify before you install)
Check each vendor's privacy policy and help docs. Summaries below are common 2026 patterns, not legal guarantees.
AICHE: Audio streamed to Groq, discarded immediately after processing, within 1 second (privacy policy). Transcripts on device by default; optional E2E encrypted sync. No training on user audio per policy. Cloud transcription; offline queue stores encrypted recordings locally until reconnect.
Wispr Flow: Cloud processing (Wispr privacy docs). Account-linked transcript history. Training per vendor policy. Cloud-only per third-party reviews; no offline queue.
Apple Dictation: On-device on supported Apple Silicon (Apple privacy). System-managed transcripts. Apple policy applies. On-device where supported.
Microsoft Voice Typing: Cloud or on-device by mode. Not a standalone note archive. Microsoft policy applies. Fluid Dictation on Copilot+ PCs can run on-device.
MacWhisper / Whisper.cpp: Local audio if configured locally. Local files you control. No cloud if fully local.
Otter.ai: Meeting audio retained per plan. Cloud account history. See Otter policy.
Dimensions to ask about: audio retention, transcript storage, training use, local vs cloud, clipboard access, active-window monitoring, wake-on-voice default, correction logging.
Voice apps are easy to install and hard to audit
You press a button, you talk, text appears. Under that simple interaction, the app may read which window is active, what is on screen, or what is on your clipboard.
Many 2026 voice apps advertise some mix of:
- Per-app adaptation (Slack vs Gmail vs IDE cleanup)
- Smart paste (tone or format tuned to paste destination)
- Hands-free start (mic open without a button)
- Learning over time (corrections and style models)
Each feature is useful. Each also has a mechanical cost. Below is what those features need, and what AICHE chose to build or skip.
How we think about privacy at AICHE
Reduce leak surface. We ask whether a feature's value is worth the data it touches. If the only path needs window monitoring or clipboard scraping, we often skip the feature.
Narrow permissions. Broad OS permissions create a broad surface to worry about. We try to keep ours narrow.
Stay honest. Audio leaves your machine for Groq, is processed in seconds, and is discarded. Transcripts stay on device by default. Opt-in sync is end-to-end encrypted with your passphrase. We are not local-only and we do not claim zero data movement.
Adapting to the app you're in
Apps like Wispr Flow describe learning your words across apps. Aqua Voice markets screen context ("Your screen is its dictionary").
To tune per app, a tool typically needs which app is active and often what is on screen, for as long as it runs.
We did not build per-app adaptation. AICHE's cleanup pipeline (filler removal, paragraph normalization, AI cleanup, custom vocabulary) runs the same regardless of focused app. We do not send active-app or screen context to the transcription model.
Smart Insert stores transient focus metadata when you press the hotkey (window and text field) so text returns where recording started. That metadata is not used for per-app output tuning.
Tradeoff: consistent output everywhere; no Slack-vs-Gmail tuning unless you edit after insert.
Pasting smart
Smart paste needs clipboard access and often surrounding document context at the paste target.
We do not monitor the clipboard. AICHE inserts at the cursor via Smart Insert. We do not read clipboard history or paste-destination content. Inserted text is what you just dictated, after cleanup.
Tradeoff: no paste-aware tone shifts. Copy a JIRA title, dictate, and AICHE will not read the ticket to tune output.
Start talking, we're listening
Hands-free and wake-on-voice modes keep the microphone analyzing audio to detect speech start. OtterPilot applies a similar idea to meetings. Wispr hands-free docs describe push-button-free capture.
AICHE default: push-button. ⌃+⌥+R (Mac) or Ctrl+Alt+R (Windows/Linux) toggles recording. Mobile: tap mic or widget. Mic is off until you start.
Voice Code (Pro) is separate, opt-in continuous listening for piping speech into AI coding agents. Off by default. A visible floating bar shows while it runs, with mute. When active, the mic stays open and speech types at the cursor. We document it here because omitting it would be inaccurate.
Tradeoff: one extra button press vs mic open whenever the app runs.
Learning over time
"Gets smarter with use" can mean shared dictionaries, per-user correction logs on vendor servers, or behavioral style models.
AICHE's only per-user learning surface is custom vocabulary: up to 50 entries you add (names, brands, acronyms, code identifiers). It syncs via the same E2E encrypted sync as notes. We do not silently log corrections or train on your audio.
Tradeoff: no evolving write-style model; you maintain vocabulary deliberately.
Permissions, briefly
AICHE does not request ongoing window monitoring for adaptation. Smart Insert reads which window and field had focus at hotkey press, then restores focus there after processing. That metadata is local and transient, not sent for model context.
We do not monitor the clipboard except brief paste simulation when Smart Insert uses clipboard mode. Mic access is when you hotkey (or when Voice Code is on). File access covers transcripts and the offline queue. Mac Accessibility permission is for keystroke simulation during insert, not for reading screen content.
Read the OS permission dialog for any voice app. If permissions exceed the stated feature, ask why.
Where AICHE does not win on privacy
Fully local processing. If audio must never leave the device, consider local Whisper tools: MacWhisper, VoiceInk, Speech Note (Flatpak), or a Whisper.cpp setup. You trade polish speed for local-only audio.
One-time purchase, no cloud account. MacWhisper offers a one-time Mac license. AICHE is subscription-only.
We name these because some users' hard constraints are real.
Questions worth asking (any voice app)
- Does it listen when I have not pressed record? Is that opt-in and visible?
- Does it read my clipboard? Only on paste, or whenever it runs?
- Does it know which app I am in? Is that sent anywhere?
- Does it read window titles?
- Does it store corrections I make? Where, and for how long?
- What audio does it keep, and where is "deleted" defined?
- If I cancel, are transcripts portable and deleted from their servers?
- What third-party SDKs ship (analytics, crash reporting)?
AICHE's answers (same eight questions)
Listening without record: Default push-button only. Voice Code opt-in, visible bar, mute control.
Clipboard: Not monitored for context. Smart Insert may briefly write transcribed text for paste simulation, then restore your prior clipboard contents.
Active app: Not used for cleanup tuning or sent for model context. Smart Insert uses transient focus metadata at hotkey press only to return text to that field.
Window titles: Not sent for transcription or per-app cleanup tuning.
Corrections: Custom vocabulary only; no silent correction logging.
Audio: Streamed to Groq, processed, discarded immediately after processing, within 1 second. No persistent audio on our servers.
Cancellation: Transcripts stay on device by default. Opt-in E2E sync deletes from our servers when you cancel.
SDKs: Desktop ships without analytics SDK. Mobile uses Firebase for ad attribution only, named in privacy policy.
If those answers do not match your bar, alternatives in this category exist.
Try AICHE
7-day free trial, no credit card. Personal $3.99/mo on annual ($4.99/mo monthly). Pro $8.33/mo on annual ($9.99/mo monthly). Available on 9 platforms + REST API (Pro).