Parleq /PAR-lek/ is a local-first AI dictation app for macOS. Hold a global hotkey, talk, and watch your words appear — cleaned up, punctuated, and pasted into whatever app was focused. Audio stays on your Mac.
Apache-2.0 · macOS 14+ · Apple Silicon recommended
Refinement
Most dictation tools clean up your words once and you're stuck with whatever they decide. Parleq lets you keep talking. While the overlay is open, hold the hotkey again — your next sentence becomes an edit instruction over the text you already have. Change tone, restructure, add context, switch formats, all without touching the keyboard. When it's right, tap the hotkey to paste.
For your inbox
You're at your desk; you want to send your manager a status update. Talk first, polish second.
You dictate
"wanted to give you a quick update on the q3 numbers we're tracking ahead of plan on revenue but a little behind on new accounts"
Parleq cleans →
"Wanted to give you a quick update on the Q3 numbers. We're tracking ahead of plan on revenue but a little behind on new accounts."
You hold ⌥ and say
"make it a proper email starting with hi cynthia, and ask for a 15-minute check-in this week"
Parleq refines →
"Hi Cynthia, I wanted to share a quick Q3 update. We're tracking ahead of plan on revenue but a little behind on new accounts. Could we set up a 15-minute check-in this week to review priorities?"
↳ tap ⌥ — pasted
For your code reviews
Mid-review, you spot a bug. Talk through what you found, tighten the wording, then reshape it for the right audience — all without typing.
You dictate
"hey can you take another look at the auth middleware pr i think there's a bug in the token refresh path where we're not handling the 401 retry correctly"
Parleq cleans →
"Hey, can you take another look at the auth middleware PR? I think there's a bug in the token refresh path where we're not handling the 401 retry correctly."
You hold ⌥ and say
"shorter, drop the hedge"
Parleq refines →
"Can you re-review the auth middleware PR? The 401 retry on token refresh is broken."
You hold ⌥ and say
"rewrite as a slack message tagging at sarah, and add that I'll add a regression test"
Parleq refines →
"@sarah can you re-review the auth middleware PR? The 401 retry on token refresh is broken — I'll add a regression test."
↳ tap ⌥ — pasted into Slack
Each refinement runs against the visible text, not from scratch. Tone changes, format swaps ("rewrite as bullets", "turn it into a commit message"), corrections, or building up a longer thought across several presses — all the same loop, all driven by your voice.
Your audio never goes to the cloud. Parleq bundles Parakeet TDT v3 — a small, fast speech model — and runs it locally on Apple Silicon's Neural Engine. No cloud transcription bill, no per-call latency, and after first launch, no warm-up wait.
Plug in whichever AI provider fits your situation: Google Gemini for personal use, AWS Bedrock (Claude or GPT-OSS) for an AWS shop, Google Vertex AI for a GCP shop, or Azure OpenAI for a Microsoft shop. Or skip cleanup entirely and paste the raw transcript. Each provider works with both pasted API keys and the same sign-in your other tools use — gcloud, az login, AWS SSO, IAM keys, or scoped Bedrock API keys.
Names, jargon, and project codenames the speech model commonly mishears go in the custom dictionary. Add the alternate spellings the model usually emits, an optional context blurb to help the AI judge whether a topic actually matches, and a per-term toggle for cases where biasing the speech recognition itself causes false positives.
The app you were in when you pressed the hotkey is captured at that moment. If your attention drifts mid-dictation, the cleaned text still lands in the right window. Per-app overrides skip the trailing space for terminals and other apps that handle their own spacing.
Audio is never written to disk — it lives in memory only, for the few seconds a dictation takes. Logs record metrics (latency, token counts) but never the text itself. Provider API keys live in the macOS Keychain. Cloud sign-in modes (AWS SSO, gcloud, az login) delegate to your existing CLI session, so Parleq never stores long-lived AWS, GCP, or Azure session tokens directly.
The last 20 cleaned dictations are kept in memory and surfaced in the menu bar — click any entry to copy it back to the clipboard. Useful when a paste lands in the wrong app. Wiped on quit; never written to disk.
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.