Key Takeaways
Understanding a voice note in a language you do not speak takes two separate operations: turning the audio into text in the language it was actually spoken, then translating that text. Your translation app cannot begin until the first one is finished, because it expects typing. The transcription step is where quality is decided, and a mistake there gets translated faithfully into your language and stops looking like a mistake. Do both steps yourself and nobody else has to hear the message.
A voice note arrives from a relative, a supplier, a landlord. You catch a few words, enough to know it matters and not enough to know what it says. The obvious move is to open the translation app you already have on your phone, and the obvious move does not work, because the app is waiting for you to type and you cannot type what you cannot hear.
The second obvious move is to forward it to someone who speaks the language. That works, and it costs you something you may not want to pay. This article is about doing it without either of those, what the machine gets right, and the specific places it goes wrong that you can learn to spot even in a language you have never read.
Why the translation app you already have cannot start
Translating speech is two operations wearing one name. The first is transcription: converting the sound into written words in the language the speaker used. The second is translation: converting those written words into a language you read. They are separate pieces of machinery, they fail in different ways, and most of the tools people reach for only do the second one.
That is the whole reason the app on your phone sits there doing nothing useful. It is built around a text box. Text goes in, text comes out. Handed a sound file, it has nothing to work with. Some tools accept speech through a microphone, which helps when you are the one talking and can repeat yourself, and helps much less with a recording of somebody else speaking quickly on a bad connection.
Once you understand the order of operations, the problem changes shape. You are not looking for a translator. You are looking for something that will produce a written version of the message in its original language, after which your existing translation habits work exactly as they always have. Whatever you already use for text is fine for the second half. The first half is the part you have to solve.
This also explains something people find puzzling: why a message can come back translated into fluent, confident, entirely wrong English. The translation half did its job. It was handed bad input and rendered it beautifully.
The cost of asking someone to listen
Forwarding the message to a bilingual friend, a cousin, a colleague who happens to speak the language, is fast and usually accurate. It also means handing that person the contents of a private conversation, in full, with tone of voice included.
For a message about a delivery date, nobody cares. For messages about money, health or family, the calculation changes. A voice note from a relative about a medical result is not something you want a colleague to hear. A landlord’s message about arrears is not something you want to explain to a friend before you have decided what you think about it yourself. Neither is a supplier’s message about a payment you may have got wrong.
What happens instead is that the message stays unopened. Not because it is unimportant, but because dealing with it means involving somebody, and involving somebody has a social price that gets paid before you even know what the message says. People leave voice notes sitting for days over exactly this. The message they cannot read is the one they least want read aloud.
That is a real reason to run the two steps yourself, and it is worth saying plainly rather than treating privacy as a bonus feature. The point is not that machine transcription is better than your cousin. Your cousin is probably better. The point is that you get to read the thing first, alone, and decide afterwards whether anyone else needs to know.
The first step decides the quality of the second
Everything about how useful the result is comes down to the transcript in the original language. Get that right and the translation is a routine problem. Get it wrong and the error travels.
Here is what makes that dangerous. If a transcript in the original language contains a mistake, and you cannot read the original language, you will never see the mistake. It goes into the translation step, gets rendered accurately into a language you do read, and arrives looking like a normal sentence. There is no jagged edge. Nothing flags it. A wrong word that survives translation looks exactly like a right one.
Compare that to a transcript in a language you speak, where an error announces itself. You read a sentence, it does not make sense, you play the audio again. That correction loop is unavailable to you here. You are reading the output of a process you cannot audit.
The practical consequence is that you should treat the transcript, not the translation, as the thing to be suspicious about. If something reads oddly in your language, the fault is very rarely in the translation half. It is nearly always a word the transcription heard wrong, and the fix is to look at the original-language text at that point and re-listen to that moment of audio, even without understanding it. Hearing that the speaker said a short word where the transcript has a long one tells you something.
Let it name the language, then check the name
A practical obstacle appears immediately: some tools ask you to pick the source language from a list before they will do anything. If you cannot identify the language with confidence, and people often cannot distinguish closely related ones by ear, you are guessing at the setting that determines the whole result.
Detection from the audio itself removes that step. The SozAI extension for WhatsApp Web works out the language from the recording rather than asking you to choose, and puts a badge on the result stating which language it decided on. That badge is the useful part, because it is how you catch a wrong guess: if it names a language you know the sender does not speak, the transcript underneath is not worth reading and you know it in one glance instead of after two paragraphs of confusion.
One thing to expect. The summary comes back in the same language the message was spoken in. That is the correct behaviour for the first step, and it means a message you cannot read is still a message you cannot read. You have converted an audio problem into a text problem, which is the entire point, but the translation step is still yours to do afterwards.
The parts most likely to be wrong are the parts that matter
Automatic transcripts are strongest on ordinary connected speech and weakest on exactly three things: personal names, place names, and numbers. A common word can be predicted from the words around it. A surname cannot. A street name cannot. A figure cannot.
Now consider what voice notes are usually about. Where to meet. How much is owed. Which day. Who said what to whom. The information you actually needed is concentrated in the parts of the transcript least likely to be right, and in a message about an address or an amount, those parts are not detail, they are the message.
So read the transcript in two passes. The first pass gives you the shape of the message: who is speaking, what it is about, whether it needs a reply today. Trust that. The second pass is only for the specifics, and there you should verify rather than trust.
- Play back the few seconds of audio around any number and listen for the digits, which are often recognisable even in a language you do not speak.
- Check names against your own contacts or previous messages rather than against the transcript.
- Compare an address to one you already have on file for that person or business.
- If an amount decides what you do next, confirm it in writing before acting on it.
Asking the sender to confirm a figure in text is not an admission that you did not understand. It is what you would do with a number heard over a bad phone line in your own language.
What this will not do for you
It will not give you the message in your language in one move. Two steps means two steps, and the transcript arrives in the language it was spoken. If you want the finished result in your own language without running the translation half separately, that is a different job than transcription alone, and translation between languages is where to look. The same applies to audio you hold as a file rather than as a WhatsApp message, a recording, a forwarded attachment, something saved off a phone, which the transcription app handles and a browser extension for WhatsApp Web does not.
It will not tell you tone. A short transcript of a long message strips out hesitation, warmth, irritation and the pause before the important sentence. If you are trying to work out how worried a relative is, the words alone will underserve you, and you should listen to the audio again with the transcript in front of you.
It will not be reliable on badly recorded audio. Wind, a speakerphone in a car, two people talking at once, a voice note recorded in a corridor. Poor input produces a confident bad transcript rather than an obvious failure, which is worse, because it looks finished.
And it will not replace your cousin when the message is genuinely ambiguous. What it does is let you find out whether the message is ambiguous before anyone else hears it. Most of the time the answer is no, and you can deal with it yourself. When the answer is yes, you now know which thirty seconds to ask about instead of handing over the whole recording.
