DEV Community

Dudko Development
Dudko Development

Posted on Originally published at ai-translator.dudko.dev

Translating a live call on a Mac: the audio routing is the hard part

Everyone who tries this starts in the same place. The models are good now, you
have a call on Thursday with someone whose English is worse than they think it
is, and you reckon the machine can bridge it. So you go looking for a virtual
audio device, you install one, and then you spend forty minutes discovering that
your setup can do exactly one of the two things you need.

That's the real problem, and it isn't the translation. A conversation has two
directions, and macOS audio routing was never designed to fork one.

Why one virtual device is never enough

Think about what has to happen on a translated call.

Their voice arrives in the call app. You need it to go into a translator, and
you need the translated version to come out of your headphones. Fine. Point the
call app's output at a virtual device, have the translator listen to that
device, play the result to your headphones.

Now you've gone deaf to everything else. Your output is the virtual device, so
the other person's untranslated voice and the meeting chime reach your ears only
through the translator. If the translator stalls, you hear nothing at all, and
you can't tell whether the call dropped or the model did.

Your voice is worse. You speak, and it needs to reach the translator, get
rendered into their language, and arrive at the call app as if it were your
microphone
. The call app only takes one input device. Give it the virtual
device and your real microphone is orphaned: something else has to pick it up
and feed it in. Give it your real microphone and the translation has nowhere to
go.

So you need two paths that don't touch:

  • inbound: call app output → translator → your headphones
  • outbound: your microphone → translator → call app input

Two virtual devices, because each one is a one-way pipe and you have two ways to
go. This is the thing nobody tells you before you start, and it's why the
single-device tutorials end with someone in a forum thread asking why they can
hear themselves twice.

The feedback trap, which you will hit

If the translated version of your voice is audible on a device that your
microphone path also listens to, the translator hears its own output,
translates that, and sends it again. Round it goes. On a speakerphone this is
instant and obvious. On headphones it's subtler: fragments of the translator's
own voice start coming back as word salad, usually just as somebody important
starts talking.

Any routing you build has to keep the two directions separate in the audio
path, not just in your head. Two devices do that. One device with clever
software on top mostly does, until something buffers.

"Can't Apple's Live Translation just do this?"

On a Mac, not for Zoom, Meet or Teams. Apple's
announcement
describes Live Translation as integrated into Messages, FaceTime and Phone, and
on calls it's for one-on-one Phone and FaceTime calls in a set of supported
languages. A third-party call app isn't on that list, so for those calls you
still need something that sits in the audio path.

What AI Translator does with this

AI Translator is a menu-bar app for macOS
that sits in exactly this gap. The reason to look at it is the plumbing rather
than the models.

The installer adds two audio drivers of its own, one per direction, signed
and notarized with the app. The site's line is "No BlackHole, no Loopback", and
that's the useful part: you aren't assembling a routing graph out of a free
loopback driver like BlackHole,
a paid router like Loopback, and some
patience.

Then the routing:

  • In the call app you pick AI Translator Speaker as the speaker. The other person's voice goes through it into the app, which plays the translation to your headphones.
  • You pick AI Translator Microphone as the mic. Your voice goes through the app and comes out of that device, in their language.

You can set those two by hand in Zoom, Meet, Teams, FaceTime, or anything else
with a speaker and microphone setting. Or there's Autopilot, which switches
system audio when you press Start and restores it when you press Stop. That
matters more than it sounds. What stops people using these setups a second time
is finding their Mac still routed through a virtual device the next morning,
on a call that doesn't need translating.

The two directions run as two separate translation sessions, and if one
stalls the other keeps going. That's the design answer to the deafness problem
above. Half a call surviving is recoverable. The whole call going quiet is not.

Picking the engine, which is the only real decision

Three options, and they differ in kind, not just in price.

On-device. Speech recognition, translation and speech all run on the Mac.
29 target languages, nothing to pay, no API key. The site's claim is that
"Audio and text never touch a network", so this is the one for a confidential
call, and the only one where privacy is a property of the architecture rather
than a promise about retention. It speaks after each phrase ends rather than
over the other person. The catch is real: the on-device engine needs macOS
26.
On Sonoma or Sequoia it isn't available to you today.

OpenAI Realtime. Speaks while the other side is still talking, and in the
speaker's own voice, which is stranger than you'd expect: the other person hears
something close to you, speaking a language you don't. 13 target languages. Your
own OpenAI key, billed per minute by OpenAI.

Gemini Live. Also speaks while the other side is talking, in a natural
generated voice, and covers 29 target languages. Your own Gemini key, billed by
Google.

Both cloud engines send call audio to that provider's API. The app's framing is
"no server of ours in the audio path", which is accurate, and it isn't the same
sentence as "nothing leaves your machine". It leaves. It goes to OpenAI or
Google under your own key, on their terms.

The practical rest: macOS 14 Sonoma or newer, an Apple Silicon Mac (M1 or
later; Intel isn't supported), wired headphones if you have them (Bluetooth adds
100–300 ms), your API key in the macOS Keychain, no analytics in the app. The
app and its licence are free; registering your email gets a key valid for six
months, and renewing is the same email again.

Where this is the wrong tool

Anything legal, medical or contractual. Use a human interpreter. Not
because the models are weak on vocabulary, but because nobody on that call is
accountable for a mistranslation, and in those settings somebody has to be.

When you need a record. This is live routing. The captions window shows
both sides as they stream, but captions are kept in memory and never written
to disk. If the deliverable is a transcript, you want a different kind of tool.

Fast overlapping speech. Three people talking over each other is still hard
for everything in this category, and a call where that's normal won't be
rescued by routing.

macOS 14 or 15, and the audio can't leave the building. The on-device
engine isn't available to you and the cloud engines break your constraint. The
honest answer is that there's nothing here for you until you upgrade.

If you just want it working before Thursday

Install it, pick Gemini or OpenAI, and let Autopilot handle the devices on the
first run rather than setting them yourself. Then test it with a friend for five
minutes before the call that matters. People who skip that step find out on the
call that their headphones were still the output when they needed to be the
input.

If you're on macOS 26 and the call is sensitive, use the on-device engine and
nothing leaves the Mac.


Originally published at https://ai-translator.dudko.dev/blog/translate-a-call-on-macos-audio-routing.

Top comments (0)