Paolo Rosson

Writing

EvenMic: making Even Realities G2 glasses a Mac microphone

How I made Even Realities G2 glasses a system-wide Mac microphone, cut full-screen images from 9.4 s to 449 ms, and what it taught me about AI on devices.

Over the summer I have spent my evenings building on the Even Realities G2, everyday display glasses with a micro-LED display in each lens, a microphone and an optional ring for input. The main result is EvenMic. It turns the glasses' microphone into a normal microphone on macOS, so any Mac app can use it, from a video call to a dictation tool, with no setup per app.

That was not a given. The G2 uses Bluetooth Low Energy only and has no Bluetooth audio profile, so a Mac cannot pair with it as a headset. EvenMic receives the audio from the glasses, decodes it on the Mac and presents it as a virtual audio device. As far as I can tell, nobody had made BLE-only glasses work as a system microphone on any operating system.

A full-screen image now reaches the lens in 449 ms, down from 9.4 s on my first working version. The lens shows my dictation as I speak, small dashboards, and where a sound is coming from. I also maintain a public list of about 305 resources for G2 developers, which has 110 stars on GitHub. The launch post on X has a short demo.

Why glasses matter for AI

Most AI assistants still sit behind a keyboard or a phone screen. Glasses change both ends of the loop: you speak, and the answer appears in your field of view while your hands stay free. Even Realities raised $150M at about a $1B valuation in July, and IDC forecasts 13.6 million smart-glasses units in 2026.

The hard limits are latency and battery. Every step in a voice pipeline adds delay, and delay is easy to notice on a display in front of your eyes. In my measurements, streaming the microphone used 8.74% of the glasses' battery per hour, against 1.27% when quiet: roughly 11 hours of use against 79. Hardware teams need AI people who count radio time and battery as part of the budget.

What I built

EvenMic. Against the MacBook's built-in microphone, same voice, the G2's noise floor sits about 14 dB lower. The near-silence between sentences made dictation apps decide I had stopped speaking, and a tiny, inaudible noise floor fixed it. The G2 is also about 5 dB quieter overall and 18 to 23 dB lower between 2 and 8 kHz, the band that carries consonants, so EvenMic offers a boost for both. The glasses' audio clock runs about 1% slow. Left alone, that drops a word roughly every 18 seconds, so EvenMic corrects for it continuously.

With the R1 ring, dictation is hands-free: tap to start, tap to stop, swipe to send. I watch my words correct themselves on the lens before they reach Claude Code. EvenMic can connect straight from the Mac with no phone in the loop, or through the phone over Wi-Fi so the glasses' own features keep working.

Full-screen images in 449 ms. My first full-screen image took 9.4 seconds. Most of that was my own code waiting longer than it needed to between sends, and fixing that brought it to 2.9 s. Compression took it to about one second. Sending half the frame through each temple arm at the same time took it to 449 ms. The glasses and their firmware did not change.

Captions and dashboards. Any script on my Mac can put a message of up to about 440 characters on the lens. When a dictation ends, the transcript flashes up on the lens. A live waveform runs next to my words at 4 frames per second, and in the test run all 121 frames arrived. Dashboard cards show the weather, my agenda, recent news and health data from my Oura ring.

Direction of arrival. The glasses can tell roughly which direction a sound comes from. I put that on the lens as a live indicator that turns as I turn my head, with no phone involved. After calibrating on me, it responds in about half a second.

The developer list

In June I started awesome-even-realities-g2, a curated list of resources for building on the G2. It has about 305 entries across 16 sections, starting with Even Realities' own developer portal, SDK documentation and starter templates. Every entry is open source or official, so a reader can clone it and learn from it. It has 110 stars so far.

EvenMic depends on other people's open-source work, including men-g2-ble-gateway for the Bluetooth connection, MentraOS, BlackHole for the virtual audio device and FluidVoice for dictation. Above all it depends on the G2, which is well-made hardware with a developer platform that welcomes this kind of experiment.

What I learned about building AI for a device

Test on the face. The direction indicator worked at my desk long before it worked on my face. Four problems only appeared while I wore the glasses.

Fit the pipeline to the microphone. Speech tools are tuned for laptop and phone microphones. A different microphone, even a quieter one, can break assumptions that nobody wrote down, such as how much noise means "still talking".

Expect silent failures. During development, a dead microphone, a dead output stream, a dropped connection and a stalled event loop each caused a silent stall. EvenMic now watches for each of them. Clock drift was the same: nothing crashed, and a word went missing every 18 seconds or so.

Treat battery as a design input. Streaming costs about 6.9 times the battery of idle, so I separated the trigger from the capture. The ring and glasses can start a dictation while the Mac's own microphone does the recording.

Keep a lab notebook. I build with Claude Code, and every experiment gets a numbered row with its result. There are more than 260 rows now, and every measurement on this page comes from one of them.

Limits and what's next

The direct path needs the phone's Bluetooth off, because the glasses hold one connection at a time. The glasses record at 16 kHz, which is fine for speech but has nothing above 8 kHz. EvenMic runs on Apple Silicon Macs only, and the direction indicator is a rough guide for now.

Next I want a hands-free loop: I speak, a model on my Mac answers, and the answer appears on the lens. EvenMic is a personal project and is not affiliated with or endorsed by Even Realities.

If you work on a hardware team and these problems look familiar, I would like to hear from you. Get in touch.