You speak
so we processed eight point nine million rows. did the build pass and is it ready to ship?
the parlay docs are live. shipped some updates to kiwi.
Parleq /PAR-lek/ is a local-first AI dictation app for macOS. Hold a hotkey and talk; the cleaned-up result pastes into whatever app was focused — speech recognition always runs on your Mac, and cleanup can too, or use the cloud you bring.
Apache-2.0 · macOS 14+ · Apple Silicon required
What Parleq does
Between the hotkey and the paste, a set of small, sharp corrections each do one thing well — punctuation, numbers, your names, a cloud rewrite when you want one. Switch families and click any one to watch it work. Cleanup runs on your Mac; refine uses the AI you choose — on-device or a cloud you bring.
Numbers & percents — Spoken numbers become digits — 8.9 million, 45%.
Before
Did the build pass and is it ready to ship
After
Did the build pass and is it ready to ship?
Before
We processed eight point nine million rows
After
We processed 8.9 million rows
Before
Shipped some updates to kiwi
After
Shipped some updates to Keavi
Before
The parlay docs are live
After
The Parleq docs are live
Before
Deploy it to the code base tonight
After
Deploy it to the codebase tonight
Before
So the Q3 numbers are tracking ahead on revenue but behind on new deals — I mean new accounts
After
The Q3 numbers are tracking ahead on revenue but behind on new accounts.
“make it a proper email to Cynthia asking for a 15-minute check-in”
Before
The Q3 numbers are tracking ahead on revenue but behind on new accounts.
After
Hi Cynthia, could we set up a 15-minute check-in this week to go over the Q3 numbers?
Before
we are ahead on revenue and a little behind on new accounts and we should sync this week
After
Ahead on revenue, slightly behind on new accounts — let’s sync this week.
Before
hey can you get me the q3 deck when you get a sec
After
Hi — could you send me the Q3 deck when you get a chance?
Dictating in Mail…
Choosing a window…
Your dictation
Before
does thursday work for the check in
After
Looks like Thursday at 2pm is open — does that work for the check-in?
Give the model context — tap the Space key in the window above.
In Parleq you'd hold your hotkey and tap Space mid-dictation. Here, just click the controls in the window.
Ready in Mail…
Transcribing on your Mac…
Raw
Cleaned
So we shipped the update to Keavi — 8.9 million rows.
Click a number to undo a correction, or .
Pasted
So we shipped the update to Keavi — 8.9 million rows.
Hold your hotkey and speak — here, tap to talk.
In Parleq this is one press-speak-release. Here, tap the controls to step through.
That dictation didn't land — the window closed before it pasted.
Re-running from memory…
Recovered
Following up on the Q3 review — can we lock the roadmap by Friday and share it with the team?
Closed it too soon? Your last dictation is still in memory — tap R.
In Parleq your last dictation's audio is held in memory — tap R to re-run it, no re-speaking.
You said
So we shipped the kiwi update — uh, no, two nights ago — up forty percent.
Fixed · Instant
So we shipped the Keavi update — uh, no, two nights ago — up 40%.
Runs on your Mac — deterministic, no network, no wait.
Same dictation — you set how much cleanup runs.
Every provider is your own account; Parleq runs no server in the path.
Skip the overlay — double-tap and hold, speak, release. Text lands right in the field.
Double-tap and release instead for hands-free continuous dictation.
Teach a sound-alike: Keavi not “kiwi”
Read aloud: “We shipped the update with Keavi.”
Listening…
Next time you dictate
let's sync with kiwi tomorrow
Heard correctly
Let's sync with Keavi tomorrow.
Sound-alikes trip up any recognizer. Teach Parleq your voice — read the line aloud.
Parleq stores a voiceprint — your acoustics, not audio or a transcript — encrypted on your Mac.
You dictated
so um can we push the sync to thursday and i'll send the deck
Draft
Can we push the sync to Thursday? I'll send the deck.
“make it a polite email to Sam”
Rewritten
Hi Sam — could we move our sync to Thursday? I'll send the deck ahead of time.
“offer Friday as a backup”
Rewritten
Hi Sam — could we move our sync to Thursday, or Friday if that's easier? I'll send the deck ahead.
One tap to finish
Tap a preset to restyle in one more pass — or just Accept.
You dictated a draft. Reshape it by voice — tap to refine.
Each hotkey press treats your speech as an edit instruction — voice undo/redo.
Per-app cleanup
Set which cleanup each app gets — instant on your Mac for a terminal, polished for email and chat. Parleq remembers, per app.
Instant — on-device Polished — the provider you set
App defaults
Privacy & control
Speech recognition runs on your Mac, and your audio never leaves it. Cleanup runs on-device too — or through your own cloud account, if you bring one. Keavi runs no server and is never in the data path.
Parakeet speech recognition runs on the Apple Neural Engine. Your audio stays in memory on your Mac — the last dictation is kept so you can re-run it — never written to disk, never sent anywhere.
Cleanup runs on-device, or through your own Gemini, OpenAI, Bedrock, Vertex, or Azure account. Your keys, your data path — Parleq just talks to the endpoint you configured.
Audio stays in memory, logs are metrics-only, secrets live in the Keychain, and dictation fails closed when you sign out. More for IT teams →
Your last dictation is still here
Held in memory — never written to disk.
Never lose a dictation. Dismissed one with a stray key, or pasted into the wrong app? Your last dictation stays in memory — hold the hotkey and tap R to re-run it. Parleq also speaks up the moment it stops hearing your mic, so you never paste silence.
For organizations
One corporate sign-in backs cloud cleanup for every employee — not just the engineers with a CLI configured. The user authenticates once with your identity provider (Okta, Microsoft Entra ID, or any OIDC IdP); Parleq federates that into AWS Bedrock or Google Cloud Vertex AI and mints short-lived per-cloud credentials at call time. No per-user API keys to distribute or rotate, and no long-lived cloud secret on the device.
First launch downloads the speech model (~150 MB). The menu-bar icon switches from a download glyph to a microphone when ready — typically 30–60 seconds on first run, under 5 seconds on subsequent launches.
Apache-2.0 · macOS 14+ · Apple Silicon required · free and open source