Inside a Pronunciation API: How 'Niamh' Finally Sounds Right

Why this matters for a voice AI agent

Everyone has had the experience of a name being mangled by an automated system, a delivery text, a call center greeting, a boarding announcement. It seems small, and it carries real weight anyway: getting someone's name right is one of the most basic signals of attention a system can give, and getting it wrong, consistently, in the same wrong way, every time, reads like carelessness, the kind that comes from nobody thinking to check.

Ireland is a useful place to look at why this keeps happening, because Irish given names are common, well documented, instantly recognizable once you've heard them said correctly, and almost impossible for an English-trained system to guess. Niamh, Saoirse, Siobhán: none of these follow English spelling-to-sound rules, and there's no clever prompt or careful phrasing that fixes it, because the failure happens before the system ever gets to "understanding" the word. It's converting spelling to sound, one letter at a time, with no idea it's looking at a name at all.

Why Voice AI Mispronounces Irish Names

Most text-to-speech systems work by predicting how a word probably sounds based on patterns learned from enormous amounts of English text. That works well for English words, and reasonably well for common loanwords. It falls apart for names drawn from a different orthography entirely. Irish (Gaeilge) spelling maps letter combinations to sounds in ways that simply don't exist in English: "mh" reads as a "v" sound, vowel clusters carry meaning English readers have no reference for, and a name that looks like it should be three syllables is often one.

The fix is telling the system, explicitly and permanently, how a specific word should sound, and having that stick. Under the hood, that's a small piece of infrastructure: a lookup table that intercepts a known word before synthesis and swaps in a pronunciation that's already been verified correct, rather than leaving it to inference every single time. Set up once, applied forever after, no special formatting required in the sentence itself.

Every example below plays the same word two ways: Default (what you get with no correction) and Corrected (with that lookup applied). Listen to both: this is real audio, straight from the API.

Common Irish Names, Correctly Pronounced

Irish given names are drawn from Irish orthography, where letter combinations map to sounds with almost no relationship to English spelling rules. A model trained mostly on English text has no statistical basis for guessing right.

Irish Phrases Beyond Names

Names are the obvious case. These are the ones that catch people off guard: official titles, place names, and tourism terms that show up constantly in ordinary conversation once you're actually in Ireland, or talking to someone who is.

More Ways Pronunciation Control Helps Voice AI

The same fix applies well beyond names too. A few other situations look unrelated at first and turn out to be the same problem wearing a different hat.

1. Alphanumeric codes

Irish Eircodes mix letters and digits, for example "D02 AF30," and a "0" read as "zero" instead of the letter "O" is a genuinely common source of confusion, the kind of small ambiguity that derails an otherwise fine conversation. Anchoring the exact pattern to a spelled-out reading removes it entirely.

2. A greeting in the middle of a sentence

Irish (Gaeilge) shows up alongside English on official signage and public services constantly. Dropping a short, correctly pronounced Irish phrase into an otherwise English sentence, without switching voices or languages entirely, is a small trick that lands as genuine familiarity rather than a translated afterthought.

3. Slang that isn't spelled the way it sounds

Informal, commonly used words that don't look like they sound are exactly the kind of input that gets flattened or mangled by systems built for careful, formal English. Getting it right is the difference between sounding like a local and sounding like a phrasebook.

The takeaway

None of this is unique to Ireland. Every language, every region, has its own version of Niamh: a name or place or phrase that's common, correct, and reliably mangled by systems that were never taught it properly. The underlying fix is the same wherever it shows up: a small, explicit correction, set once, applied every time after.

That's the kind of problem Telnyx spends a lot of time on. Voice AI is only as good as the details it gets right consistently, and pronunciation is one of the details people notice immediately even when they can't articulate why something felt off. Telnyx's Pronunciation Dictionaries API is built for exactly this: register a correction once, and it applies automatically from then on, across an entire voice agent's call flow, no per-call workarounds. It's a small piece of infrastructure, but it's the kind that determines whether a voice AI system feels like it was built with a place and its people in mind, or built somewhere else and shipped everywhere. If you're building voice AI and want to see how the rest of the stack handles problems like this one, the Telnyx TTS API is a reasonable place to start.