Hebrew TTS without niqqud: why words are mispronounced and how to fix it

The gender of "you"

The guess that matters most on a call is gender. Hebrew marks it even in "you": לך is lekha (לְךָ) to a man and lakh (לָךְ) to a woman, and unpointed text shows no difference. A voice agent that reads every לך as lekha addresses half its callers as men, and callers notice at once.

Gender also appears in verbs and adjectives, including in the first person in some tenses, so the problem is not limited to one word. It is a writing problem as much as a pronunciation one, because the text itself has to choose a form before the voice can say it.

Homographs

Homographs are words spelled alike but pronounced differently. אוכל is okhel (food, or eating) and ukhal (I will be able), and דבר is davar (thing) or diber (spoke). In a reply such as "לא אוכל לאשר" (I won't be able to confirm), the wrong reading turns the sentence into nonsense, and a short reply gives the engine little context to choose from.

Names and loanwords are a related trap. Words the model rarely saw in training, such as a new brand or an unusual surname, have no context to resolve them at all.

Stress and the vocal shva

Stress is the guess that niqqud does not even solve. The Phonikud project, an open Hebrew pronunciation tool, adds its own marks for stress and for the vocal shva, because standard niqqud cannot tell some words apart (Phonikud). The shva is a vowel sign that is sometimes pronounced as a short e and sometimes silent. Phonikud's example is אורז, which is either rice or packing depending on which syllable is stressed.

This also explains why some Hebrew voices sound formal or old-fashioned. Phonikud's documentation notes that the spoken modern shva differs from the shva taught in grammar books, and a voice that follows every textbook rule can sound like a reading lesson on a phone call.

Models that learn readings from context

Roth, Turetzky and Adi (Interspeech 2024) trained a Hebrew TTS model as a language model over speech units, fed with ordinary word pieces and trained on "in-the-wild weakly supervised data" (arXiv 2407.12206). They report that it beat the diacritic-based baselines they tested on both content and naturalness. The model learns readings from context, much as a human reader does, so it never needs niqqud at all.

Phonikud: spelling out the sounds

Phonikud converts Hebrew text to phonetic symbols in the International Phonetic Alphabet (IPA) for TTS, adding niqqud, stress and vocal shva on the way. Its paper, accepted to Interspeech 2026, reports that it predicts Hebrew phonemes more accurately than earlier methods, and that small local TTS models using it approach large proprietary systems (Kolani and others). The code is released under CC BY 4.0.

Both are research results, not tested phone-call products. Phonikud's own documentation lists formal-sounding niqqud, inconsistent stress and wrong predictions for names and non-Hebrew words as known limits. Its public roadmap from June 2025 asks for a better TTS model and much more hand-corrected data (issue 111). Treat these tools as components to evaluate, not as a guarantee.

Testing before launch

Test the sentences callers will actually hear, not sample text. Export the 50 or so replies your agent says most often, including names, amounts and dates, and have a native speaker flag every gendered "you" and every word that could be read two ways. For each one, decide whether to rephrase it or point the word.

Then render the flagged sentences and have a native Israeli speaker listen on a real phone line, checking vowels, stress and gender. Repeat after any change to the prompt or the voice, since a new model version can resolve the same word differently. Our guide to testing a multilingual voice agent covers the wider process.

Hebrew and Arabic share the problem

Builders serving both Israel and Arab markets meet the same problem twice. Arabic is also written without most short vowels, and the same rules apply: rephrase, mark sparingly and test with native listeners, as our guide to Arabic TTS and tashkeel explains. Hebrew adds unmarked stress even in pointed text, while Arabic adds dialects that change the words themselves, and its numbers and dates have their own traps (Arabic numbers, prices and dates in TTS). Persian has a version of the same issue too (Farsi text to speech).