Does AIspeaky work fully offline?

No. During the beta, recognition, correction, and translation run through managed cloud providers via a blind relay. The app captures audio locally and streams it to the provider; only technical metadata is retained.

What leaves my device?

Audio for dictation and translation, and text for post-correction and rewrite commands. Nothing else is uploaded — no clipboard, no window titles, no file contents.

Is my voice used to train models?

No. AIspeaky uses provider endpoints configured with training opt-out. Payloads are processed for the request and discarded on their side per each provider's retention policy.

Do you store my transcripts?

No content is stored on AIspeaky servers. Only anonymous technical metadata (latency, error codes, provider region) is retained for diagnostics.

Which regions are used?

Requests are routed to the closest supported region for each managed provider. The full region matrix is listed on the security page.

Privacy & security

Honest about the cloud.Quiet about your content.

AIspeaky uses managed providers for the strongest recognition and translation, hidden behind a first-party relay. This page is the map: what leaves the device, why, where it goes and how long it lives.

Download beta Back to product

AIspeaky is currently in beta and free to use within beta limits. We do not read or store the content of what you dictate.

At a glance

Every managed surface, one table.

Same shape as the home-page summary — expanded with per-surface deep dives below.

Route

What leaves the device

Purpose · retention

Status

Speech recognition
Raw audio buffer of the current utterance
Turn your voice into a transcript we can insert. Audio is streamed for one recognition request at a time.
AIspeaky does not persist the audio or the returned transcript after the response is delivered to the client.
Managed · in beta
Realtime translation
The recognized source-language transcript
Return a target-language rendering of the source line so you can insert it into the focused field.
AIspeaky does not persist source or target text after the response is returned.
Managed · where enabled
Voice commands
For simple commands: nothing beyond the app-level metadata already sent for diagnostics
Execute deterministic shortcuts locally, or ask a managed model to edit a selection when you use a rewrite command.
Simple commands never leave the device.
Simple set · deterministic
Beta diagnostics
Anonymous device and OS fingerprint (version, arch, locale)
Detect regressions before you report them; keep beta usage inside the free tier limits.
Retained for the duration of the beta program.
Always on during beta

AIspeaky collects technical diagnostics and usage metadata during the beta. Content handling depends on the workflow — review the details before using managed features.Full security page

01 · Dictation

Speech recognition

You hold push-to-talk, we capture the utterance, a managed speech-to-text provider returns the transcript, and AIspeaky inserts the text into the focused field.

Managed · in beta

What leaves the device

Raw audio buffer of the current utterance
Session id used to correlate crash reports

Purpose

Turn your voice into a transcript we can insert. Audio is streamed for one recognition request at a time.

Retention

AIspeaky does not persist the audio or the returned transcript after the response is delivered to the client. Provider-side retention is governed by the provider's policy — see the provider list below.

Provider

Managed speech-to-text via AIspeaky's relay

Region

EU/US routing during the beta; provider chooses the nearest inference region.

1
Client → Relay
Audio chunks over TLS; no third-party provider keys on the client.
2
Relay → Provider
AIspeaky signs the request with its own credentials and forwards audio.
3
Provider → Relay
Returns the recognized text; the relay does not log the transcript body.
4
Relay → Client
Transcript delivered back over the same session. Nothing persisted server-side.

02 · Translation

Realtime translation

Recognized text is sent to a managed translation model. Source and target are visible in the tray before you insert the target.

Managed · where enabled

What leaves the device

The recognized source-language transcript
Language-pair identifiers (e.g. en→es)

Purpose

Return a target-language rendering of the source line so you can insert it into the focused field.

Retention

AIspeaky does not persist source or target text after the response is returned. Provider retention applies to the request body per their published policy.

Provider

Managed translation via AIspeaky's relay

Region

EU/US routing during the beta.

1
Client → Relay
Source text and language pair over TLS.
2
Relay → Provider
Signed request from AIspeaky's server, no client-side keys.
3
Provider → Relay
Translated line; body not logged by AIspeaky.
4
Relay → Client
Target text and quality tag; nothing persisted on our side.

03 · Commands

Voice commands

Simple commands (undo, copy, paste, select…) resolve locally on the client. Rewrite commands like 'make this shorter' call a managed model and are shown as a separate, expanding group.

Simple set · deterministic

What leaves the device

For simple commands: nothing beyond the app-level metadata already sent for diagnostics
For rewrite commands: the selected text and the command phrase

Purpose

Execute deterministic shortcuts locally, or ask a managed model to edit a selection when you use a rewrite command.

Retention

Simple commands never leave the device. Rewrite command inputs follow the same non-retention posture as translation.

Provider

Local for simple commands · managed model for rewrites

Region

n/a for local commands; EU/US routing for managed rewrites.

1
Client (local)
Simple commands resolve on-device via the accessibility bridge.
2
Client → Relay
Rewrite commands send selection + phrase over TLS.
3
Relay → Provider
Signed request; the relay does not log the selection.
4
Provider → Client
Edited text returned; original selection replaced only after your accept.

04 · Diagnostics

Beta diagnostics

AIspeaky sends technical diagnostics that help us keep the beta stable — usage counters, crash traces, session metadata. Not the content you dictate.

Always on during beta

What leaves the device

Anonymous device and OS fingerprint (version, arch, locale)
Usage counters (sessions started, recognition seconds, error codes)
Crash traces and unhandled exceptions

Purpose

Detect regressions before you report them; keep beta usage inside the free tier limits.

Retention

Retained for the duration of the beta program. On beta exit we intend to reduce this to short-lived rolling logs.

Provider

AIspeaky (first-party)

Region

EU (primary).

1
Client → AIspeaky
Batched telemetry over TLS; no dictation content included.
2
AIspeaky storage
Indexed by anonymous device id; joinable to your account only after you opt into bug reports with logs.

Providers

Who we route through — and what they see.

AIspeaky uses managed providers for the parts local models don't do well yet. Provider identities are stable for the beta and change only with a status-page announcement.

Provider	Role	Region
Managed speech-to-text	Recognition	EU/US
Managed translation	Translation	EU/US
Managed LLM (rewrite commands)	Selection rewrite	EU/US
AIspeaky relay	Signs and forwards requests; no third-party keys on the client	EU

Provider names are intentionally generic during closed beta. Specific vendors will be disclosed on this page before the first public GA.

Your rights during beta

What you can actually do today.

Delete your account and data

Sign in, open account settings, remove your account. Beta account data is deleted; diagnostics tied to your account id are purged on the next rotation.

See what we have

Ask via the contact page. During beta we return diagnostics tied to your account within 30 days.

Opt out of managed features

Disable managed recognition and translation in settings. AIspeaky then declines those features rather than falling back to a weaker local model — that keeps the honesty of what's actually running.

Corrections

What we used to say vs. what's actually true

Earlier copy on this site overclaimed local-only behavior. Here's the honest version.

We used to say	What's actually true
Audio stays on your device.	Managed features send audio or text to cloud providers through AIspeaky's relay.
Fully private.	AIspeaky holds provider credentials so your client doesn't. Provider-side handling still applies.
No telemetry.	Beta diagnostics are on by default so we can fix issues before launch.

AIspeaky doesn't claim end-to-end encryption for managed features, and doesn't claim independent certification. If you need either before installing, use the contact page — we'll be straight about what's verified today.

Ready to try it

Install the beta with the map in your hand.

You've seen exactly what leaves the device and why. If that matches what you need, grab the Fedora or Windows beta. macOS is on the waitlist.

Download beta Ask before installing