How it works.

No black box. Here's the path both voices take — the customer's and your agent's — what happens when it fails, and what your agent sees while it's running.

  1. 1

    The call arrives on your platform

    Nothing changes about how the call reaches your agent. Your IVR, your routing, your queue logic all work exactly as they do now. Taiwa doesn't sit in front of the call and doesn't touch how it's distributed. It waits until the call is connected.

  2. 2

    The extension reads both sides, in the browser

    Your agent's call console is a web page, and the live call's audio is right there in it. Taiwa reads it in the page itself — the customer's audio coming in, and your agent's going out. Nothing is uploaded to us. The capture happens on the agent's machine and stays there.

  3. 3

    Audio streams to your translation supplier, and comes back translated

    From that browser, audio streams to the engine you chose — DeepL, OpenAI or Google Gemini — over an encrypted connection that's authorised only for that one call. Translated speech comes back the same way — and the transcript arrives with it.

  4. 4

    Both people hear each other

    Your agent hears the customer in your agent's language. The customer hears your agent in theirs. The transcript appears in the extension as it goes, lives in the browser, and ends with the session.

Where the audio goes, in detail →

What the agent sees

The popup. Agent language, customer language, and whether it's live.
The live transcript, both sides, as it happens.
Detected language, and the switch to turn it off mid-call.

Not in a browser?

Teams, Zoom and softphones are applications, not web pages, so the extension can't reach their audio. The Windows desktop app does the same job one level down, at the microphone and the speaker.

The desktop app →

It's machine translation. Sometimes it's wrong.

So the question isn't whether it fails, it's what your agent can do when it does. They see the live transcript of both sides, so a mistranslation is visible rather than silent. They see the detected language, so a wrong guess is obvious. And they can turn translation off mid-call without dropping the call — the call carries on, untranslated, and they handle it the way they would have before Taiwa existed.

For medical, legal, safeguarding and financial-advice calls, use a certified human interpreter.

What it costs the call

What you gain

The call finishes on first contact. No transfer to hunt for a language, no interpreter dialled in and charged by the minute, no callback. And no language-based routing rules to maintain once they're gone.

What it adds

A short delay while speech is translated — quicker than a human interpreter would be. Your agent talks, there's a pause, the customer hears it. Calls run a little longer than a same-language call, and far shorter than a transfer or a dialled-in interpreter.

It also covers what a small language team can't: breaks, holidays, out-of-hours, and the languages you get one call a week in — without staffing for every one of them.

Setup

  1. 1

    Install the extension from the Chrome Web Store, or the desktop app on Windows.

  2. 2

    Sign in. Your admin loads your supplier key and sets the languages and the limits.

  3. 3

    Take a call.

No change to your call platform. No migration. No professional services engagement.

Want to see it on your platform?

Request early access — approved within 24 hours. Tell us what you run and which languages you're missing.