Where Taiwa fits.
It's very good at some calls and mediocre at others, and the difference isn't subtle. Point it at the right queue and it pays for itself. Point it at the wrong one and you'll be disappointed in a way we could have warned you about.
Where it works
Transactional calls. The ones where the words carry the meaning.
Where's my order · Change my address · Rebook this · Reset my password · Is this in stock · Track my delivery · Confirm my appointment · Return this item
These calls are an exchange of information. The customer needs a fact, or needs you to change one. What matters is that the postcode is right and the date is correct — not the melody of how it was said. Translation carries facts extremely well.
This is most of a service desk's volume, and it's the volume that's currently being transferred, held, or called back.
Where it's harder
Calls where how it's said matters as much as what is said.
A complaint that's escalating · A customer who's upset · A negotiation · Anything where you're reading the room
Here the meaning isn't only in the words. It's in the pause before the answer, the edge in the voice, the sentence that trails off. Taiwa translates the words accurately and delivers them in a synthesised voice that is calm regardless of what it's carrying.
A furious customer can arrive sounding composed. That's a real problem.
Medical, legal, safeguarding and financial-advice calls aren't on this page at all. Those need a certified human interpreter.
Where Taiwa stops →What we don't translate
We translate words. Not the way you say them.
How something is said carries meaning the words don't — intonation, stress, the rising tone that turns a statement into a question. "I didn't say that" means four different things depending on which word you lean on, and translation flattens all four into the same sentence.
Emotion, sarcasm, hesitation and urgency ride on the same channel. None of it survives being turned into text and read back out by a synthesiser.
This is on DeepL's roadmap, and it's coming. In the meantime you don't lose those cues — the original voice keeps playing underneath (below).
So you still hear them
The original voice doesn't go away.
The obvious way to build this is to mute the customer and play a translation over the top. We don't. Their real voice keeps playing underneath the translation, quietly — 20% by default, and you can move it.
So your agent hears the translated words and the person saying them. The frustration, the hesitation, the relief, the speed they're talking at. Everything above that we can't translate is still arriving on the original channel, because we never took it away.
It works both directions. Your customer hears your agent's actual voice under the translation too — the warmth, the apology, the care that no synthesiser reproduces.
We don't translate the emotion (yet). We also don't delete it.
The other constraint
The other thing that shapes fit is the delay. Fast back-and-forth pays the lag on every exchange; calls where each side speaks in longer turns barely notice it.
Why there's a delay, and what we do about it →If you're trialling this, start here
Point it at your highest-volume transactional queue in a language you don't cover — order status, delivery, account changes. That's where the arithmetic is best and the limitations matter least.
Don't start with complaints. It'll work, but you'll be judging the hardest case first, and you'll conclude something about the tool that's really about the queue.
What is this costing you today? →Not sure which queue?
Tell us what your calls are actually about when you request access. If they're the wrong kind, we'll say so.