Building Klear: Quiet Calls from a Loud House

My house doesn't care about my calendar, so I built a Mac app that removes background noise from every call, in every app.

I work from home, and home does not care about my calendar.

Every important call I take comes with a soundtrack. The dogs losing their minds at a delivery truck. My son discovering something hilarious in the next room. The dishwasher hitting its dramatic third act right as I get to the part that matters. Sound familiar?

For a while I did what everyone does. I rode the mute button like a game-show buzzer, apologized mid-sentence, and took the big calls from the car. Zoom's built-in noise suppression helped... on Zoom. Then the next call was in Meet, or Teams, or whatever browser widget a client's scheduling tool spat out, and I was back to apologizing for my own house.

So I built Klear.

One Mic, Every App

Klear is a macOS menu bar app that strips the background noise out of your microphone and hands the cleaned-up audio to your calls as a new mic called "Klear Mic." You pick it once in Zoom, Teams, Slack, Google Meet, FaceTime, Discord, or anything else that takes a microphone, and that's it. Every app gets the quiet version of you.

Here's the thing: the noise is gone before any app ever sees your audio. It doesn't matter whose built-in suppression is good this quarter, or which platform the other side insists on. The dog barks. The call hears nothing.

It adds about 30 milliseconds of latency, which is a fancy way of saying nobody on the call will ever know it's there.

There's also a strength dial for when your room is more "gentle hum" than "daycare," and a few voice effects for when a meeting could use some levity.

Who This Is Actually For

Honestly? Probably not most of you. If your calls are standups and pairing sessions, a barking dog is a running joke, not a problem. Nobody ever lost a sprint to background noise.

Klear is for people whose voice is the work:

  • Salespeople doing six external calls a day.
  • Recruiters screening candidates between school pickups.
  • Consultants, coaches, and customer success folks whose calls have outcomes attached.
  • Therapists doing telehealth from a house that doesn't observe business hours.

None of them get to pick the platform, either. The person on the other end already did.

For those folks the math is simple: one flubbed first impression costs more than the app does. If that's you (or your partner, or that friend who apologizes for their dog on every call), this is the thing I wish I could have handed you a year ago.

Your Calls Never Leave Your Mac

Klear also transcribes your calls. Both sides, labeled "You" and "Them," saved as plain markdown in your Documents folder.

No bot joins your meeting. No audio gets uploaded. There isn't a server anywhere in the audio or transcription path. The speech models run on your Mac's Neural Engine, and the notes land on your disk and nowhere else.

If you've ever wanted what Otter or Granola do, but your company (or your conscience, or your client's NDA) won't let call audio anywhere near a cloud, this is the version you're allowed to use. And when the call ends, you've got a text file you actually own. Grep it, paste it into your CRM, feed it to whatever you like. It's just markdown, and you all know how I feel about markdown.

Okay, Now the Nerdy Part

If you scrolled past the sales pitch to get here, hi. You're my people. Here's what's actually going on under the hood.

The virtual mic is a Core Audio HAL plug-in. "Klear Mic" is an AudioServerPlugIn loopback device that lives inside coreaudiod. Writing one of these is a trust exercise. Your code runs in the process that owns audio for the entire machine, so the driver is deliberately dumb: no allocation, no locks, no Objective-C on the IO path. It also fails soft, because a crash in there doesn't take down your app. It takes down sound.

Capture skips AVAudioEngine. The app reads your real mic with raw AUHAL units, because AVAudioEngine's ~100 ms tap buffers would eat the entire latency budget on their own. Audio moves through C ring buffers (single-producer, single-consumer, no Swift runtime anywhere near the IO thread) into the denoiser and back out to the virtual device. On the way out, a varispeed read soaks up both sample-rate conversion and the clock drift between your mic and the driver.

Denoising is DeepFilterNet. It runs on a worker thread, because its inference allocates per frame, and that disqualifies it from the realtime thread. RNNoise sits inline as the fallback.

Transcription is Parakeet TDT on the Neural Engine.

End to end, the mic-to-meeting path adds about 30 ms.

There's a much longer post hiding in all of this. The fun part was learning how much of "real-time audio on macOS" is folklore you only find in old mailing list archives. If you want that one, let me know in the comments.

Want to Try It?

Klear runs on macOS 14+ on Apple silicon, with a 14-day free trial. No account, no card. Grab it from the Klear page under Apps, pick "Klear Mic" in your next meeting, and let the dog say whatever it wants.

And if you know someone who apologizes for their house on every call, send them the link. This one's for them.

No comments yet, add yours here.