Why Greek breaks voice AI — and what to demand

One word, ten shapes

Greek inflects. The same service is «καθαρισμός», «καθαρισμό» or «καθαρισμού» depending on where it sits in the sentence; the same surname is heard as «Πετρίδης», «Πετρίδη» or «κύριε Πετρίδη». In English the word stands still. In Greek its ending shifts in almost every sentence.

That hurts in two places. First in understanding: a system matching the word «καθαρισμός» against your list of services will not recognise it when the customer says «θα ήθελα για καθαρισμό». Then in speaking: if the reply comes out in the wrong case, a Greek ear catches it instantly — the way you would catch someone who learned English last month. Add the accent mark, which changes the word on its own: «πότε» (when) is not «ποτέ» (never); «άλλα» (other things) is not «αλλά» (but).

The number that is spoken in words

Ask a Greek for their phone number and you will not hear digits. You will hear words, in groups they choose themselves: «εξήντα εννιά, ογδόντα επτά…» — two at a time, then three together, with a small pause in between (an illustrative example; everyone breaks it up differently).

For a voice system that is two separate problems. The first: recognition returns words, and often half-converted ones — «εξήντα εννέα» (sixty-nine) can be written down as «60 εννέα», which read literally gives 60 and then 9, not 69. The second: the pauses between groups. A system that treats the first pause as "the caller has finished" cuts them off in the middle of their own number.

This is not a detail. One wrong digit means an appointment you cannot confirm and a customer you cannot reach again. It was among the first things we solved, and we keep tests that catch it if it ever breaks again — because this is the point where a whole call is lost.

The question that sounds like a statement

In English a question is visible in the word order: is Thursday OK versus Thursday is OK. In Greek the order does not change. «Σας βολεύει η Πέμπτη;» and «Σας βολεύει η Πέμπτη.» contain exactly the same words in exactly the same order. Only intonation separates them: the voice rises at the end.

A synthetic voice that reads flat turns every question into an announcement. The caller does not realise they were asked something, stays silent or answers something else, and the call stalls in an awkward pause that is not their fault. Working on Greek we found something subtler still: whether a question sounds like a question can hinge on the stress of the final word — the rise needs a syllable to land on. Which means the fix is not a setting; it is how the sentence is phrased.

The formal plural, and the greetings a Greek expects

Greek carries politeness inside the verb: «θέλετε» or «θέλεις», «πείτε μου» or «πες μου». A business phone speaks in the formal plural — and has to hold it in every sentence, to the end of the call. One «πες μου το όνομά σου» in the middle and the caller notices; to an older customer it reads as disrespect.

The same goes for the things nobody writes into a specification: «Καλημέρα» until midday and «Καλησπέρα» after it — a «Καλημέρα» at eight in the evening lands wrong immediately. A «Καλή σας συνέχεια» on the way out. Small things; also the entire first impression.

Then there are the accents. Northern Greek swallows unstressed vowels, Cretan colours its consonants differently, and the same word arrives altered. A product tested only on clean Athenian articulation loses accuracy exactly where you need it: on a real customer, from a mobile, in noise.

Listen to a real call

The fastest check is your own ear. Below is a real recorded call with Parousa, in Greek, from the opening «Καλησπέρα» to the booked appointment.

What to demand from any Greek voice product

You do not need linguistics to test a product. You need ten minutes. These tests apply to every provider — including us.

• 1. Listen to a whole call. Not a clip and not a sales video: one call from the first word to the goodbye. Greek problems surface in the flow, not in the opening sentence.
• 2. Dictate a phone number to it. The way your customer would: in words, in groups, with pauses. See whether it interrupts you halfway, and ask it to read the number back. Also ask what happens when it does not catch it — is there a second way to give the number?
• 3. Make it ask you something. Steer the conversation to a yes/no question and listen to the end of the sentence. If the voice does not rise, your customer will not know they are supposed to answer.
• 4. Check that it holds the formal plural. Talk to it for two or three minutes. The informal singular usually appears late, once the conversation leaves the script.
• 5. Speak to it the way your customers do. Half a sentence, a change of mind mid-way, a service called by the name your regulars use for it. Demo scenarios are always tidy; your business is not.
• 6. Ask whether it introduces itself as an AI. The European rules on artificial intelligence require the caller to know they are speaking to a system. A provider who hides it exposes you.
• 7. Ask where the call data is stored. GDPR applies here too: place of storage, retention period, purpose. Ask for a specific answer, not "in the cloud".

If your business runs on the phone, work out first what the calls you never answer are costing you: the missed-calls calculator runs on your own numbers rather than on averages. And if you want the basics first — what a solution like this actually does — see what an AI receptionist is and what an AI phone assistant does.

Where we stand

Parousa is built in Greek first, not translated into it: every change to the voice or the conversation goes through Greek testing before it reaches a call. The problems on this page are not theory — they are what we ran into while building it, and what we fixed.

To be equally clear: Parousa is not yet generally available — the first places will be offered gradually through the waitlist. Call data is stored in the European Union, and Parousa always introduces itself as a digital assistant at the start of the call. If you want to hear it handle your own business, leave us your email.