People expect machine translation to fail unpredictably. It does not. It fails in a small number of specific places, and once you know them, most of the damage is preventable — often without spending anything.
It fails confidently, which is the real problem
A translation system never hesitates. When it does not understand, it produces a fluent, grammatical, entirely wrong sentence. There is no wobble in the voice to warn the listener, no "unclear" marker, nothing.
This is why learning the failure modes matters more than it would for a human interpreter. A human says "sorry, I missed that". A model says something confident instead.
Failure 1: proper names become ordinary words
Names of people and places are often ordinary words in some language. Left to guess, a system will sometimes translate a name rather than keep it.
This one is measurable, and the measurement is unusually clean. We took identical audio, sent exactly the same bytes to the engine twice — once with a prepared terminology list and once without — and counted correct terms.
| Condition | Without the list | With the list |
|---|---|---|
| English, noisy hall and fast speaker | 46 of 129 | 67 of 129 |
| Greek, mild noise | 6 of 10 | 10 of 10 |
| English, clean audio and clear speaker | 48 of 48 | 48 of 48 |
Two things stand out. There were zero regressions — preparing the vocabulary never made anything worse. And in clean audio it made no difference at all.
That last row is the important one. Preparation is not a general accuracy boost; it is insurance that pays out precisely when conditions are bad. A quiet hall with a slow speaker does not need it. A full hall with a fast speaker does, and that is most of the services worth translating.
Failure 2: idioms
Figurative language translated literally is the classic failure, and the most dangerous, because the output looks fine. Some idioms travel between languages; most do not. The result is not gibberish — which would at least be visible — but a sensible sentence about something else entirely.
There is no technical fix for this. The fix is on the speaker's side and takes one second: if a point rests on an image, say it once plainly as well. "We should bear fruit — that is, our lives should show a change."
Failure 3: quoted passages
A read passage is the hardest few seconds in the room. It is dense, it uses older constructions, and it usually arrives without warning. It is simultaneously the most difficult thing to translate and the part your listeners care most about getting right.
Announcing the reference first — book and chapter, before the words — measurably improves what comes out, for the same reason the terminology list does: the system stops guessing from nothing.
Failure 4: the script problem, which is invisible until you look
This one we found by accident, and it is the most instructive of the four because the model was not actually wrong.
A speaker said, in Greek, "Μπορείς να μιλάς αγγλικά". The engine heard it correctly — and wrote it in Latin letters: Boris na Mila Sanglica, labelled as Spanish. "Αύριο" came back as Ábrio, labelled Portuguese.
What happened next is the damaging part. The pipeline received Latin text, concluded it was English, and translated it from English. The audience received guaranteed nonsense produced from audio the model had understood perfectly.
The fix turned out to be a single instruction to the engine: write each language in its own script, never romanise. On eighteen Greek windows captured from a real broadcast, correct output went from 6 to 17.
We also tried the intuitive fix first, and it was worse. Telling the model which languages to expect — a single "the speaker is talking Persian" hint — turned سلام into Devanagari. Narrowing the guess space with one anchor made the model more confident and more wrong. Passing the session's full language list worked; a single anchor actively harmed it. That is not something you would predict from the documentation.
Failure 5, which is not a translation failure at all
Worth naming because it is so often misdiagnosed: when the audio is too quiet, several widely used engines invent sentences rather than return nothing. Eight windows of non-speech audio, submitted with the language set to Greek, produced eight fabricated sentences — Japanese, Korean, Hindi, Arabic, Italian, Swedish, Romanian.
If your transcripts contain confident sentences nobody said, in languages nobody spoke, the translation is not broken. The microphone is.
Where it is genuinely strong
Setting expectations honestly means saying this too. Plain explanatory speech — the bulk of any talk — comes out well, and improved sharply over the last few years. The ordinary paragraphs are fine. It is the ornamental and the quoted language that needs care, and that is a much smaller share of a service than people assume.
What to actually do
- Give the system your vocabulary in advance rather than letting it guess. Measured worth: 46 → 67 in hard conditions, 6 → 10 in Greek, zero regressions.
- Name a reference before reading it. Free, one second.
- Say figurative points plainly once as well. Free.
- Check the microphone level before blaming the model. Most "bad AI" is bad audio.
- Have one bilingual member read along occasionally and tell you what they saw. Nothing replaces this.
- Put dates, times and numbers on a screen, not only in the air. They are trivially checkable and consequential when wrong.
A system that has been taught your vocabulary
InMyTongue loads the terminology of your setting before the service and keeps every language in its own script, in each source language it supports. New accounts include free credit. See how it handles your material.