HearMe

Dictation that has actually heard you.

Hold a key, talk, release — your words land at the cursor, in any app. Every model runs on your machine, and HearMe measures which one truly hears your voice, your microphone, your accent.

Read the source

Free, MIT-licensed. Windows & Linux builds are new — field reports welcome.

Meet Kwame at the Accra office at half past three.

hotkey held whisper turbo · on-device network connections: 0

The leaderboard said we'd be fine.

Speech models are graded on American and European voices. Tested on West-African-accented English, the chart-topper didn't error, didn't warn — it just silently dropped six words in every ten. Another model, further down the same chart, was near-verbatim.

That's not one laptop's bad luck. Stanford measured nearly double the error rate for Black speakers across five major vendors — the headline numbers are earned on U.S. benchmark audio, not on voices like these.

You can't see that on a leaderboard. You can only hear it — on your own voice.

So the app runs the test for you.

Voice setup, built in: read three short phrases once — the third is made of your own names and jargon — and every model on your machine is scored on accuracy and speed against your recording. One click applies the winner. As far as we can find, no other dictation app measures this — the rest hand you a model menu and wish you luck.

Illustrative run on Ghanaian-accented English. Your ranking will differ — that's the entire point.

Everything after transcription is local too.

Command mode ctrl ⇧ space

Highlight a rough paragraph, hold the key, say "make this formal." A local model (Ollama) rewrites it in place. Nothing sent, nothing retained.

Live typing

Words appear while you're still speaking, then settle into the cleaned-up text when you release. Experimental, macOS-only for now, and honest about it.

Custom words

Kwame, Adjoa, cedis — names and jargon spelled right, with one-click suggestions mined from your own dictation history.

No mockups — this is the app.

HearMe Models page: Whisper Turbo active, marked best for accented English — winner of the live accent A/B on this machine
Every model carries a verdict earned on this machine's voice — not copied from a leaderboard.
Voice setup wizard: read three phrases aloud and every selected model is scored on your voice, your microphone, and your accent
Voice setup: three phrases in, one recommended model out.
General settings: transcribe shortcut, push to talk, spoken punctuation, language and microphone
Push to talk, spoken punctuation — and your accent color on every control.
Advanced settings: overlay, model unloading, paste method and clipboard handling
The knobs go deep: overlay, model unloading, paste method, clipboard handling.

Don't trust us. Ask your machine.

Cloud providers aren't disabled — they're removed from the code, and scrubbed from old settings on load. The only network use is the model download you trigger once, checksum-pinned. Run this while dictating (macOS/Linux — on Windows, watch the app in Resource Monitor's network tab):

$ lsof -i -a -p $(pgrep -x handy) (no output — zero open connections)

Want something built like this?

HearMe went from fork to shipped, adversarially-reviewed product by one developer pairing with Claude Code — spec, code, reviews, docs, this page. I take on a small number of client builds: internal tools, automations, software that has to keep your data yours. Open a project inquiry.

Fuel the next feature.

HearMe is free and stays free. If it replaces a subscription for you, a coffee-sized tip says thanks — by card or bank via Wise, or in Bitcoin. And if you'd rather not spend a thing: a star helps just as much — it's how the next person finds this.

Star on GitHub Tip via Wise
Bitcoin — SegWit bc1qnjhgudw7837muxda75jarz2nd2wnztrzfh5mzz