ElevenLabs Hebrew: What Eleven v4 Is and How to Get an Israeli Voice That Sounds Real

For years Hebrew was a second-class language in text-to-speech: voices with an American accent, numbers read digit by digit and names that fell apart. With Eleven v3 and then v4, ElevenLabs became the main supplier of Hebrew voices for AI agents, and anyone building or buying a voice agent in Israel should understand what these models do and where they still fall short.

What Eleven v4 is

Eleven v4 is ElevenLabs' main text-to-speech model, and according to the official documentation it is the company's most emotive, highest-quality model. Three numbers matter to anyone building in Hebrew.

• More than 90 supported languages, with Hebrew (heb) on the official list.
• Up to 10,000 characters per request, enough for any voice-agent reply and most recorded messages.
• Multi-speaker dialogue through the Text to Dialogue API, useful for recording sample conversations.

Alongside v4 there is Eleven v4 Turbo: the same family, the same 90+ languages, built for real time with a median latency of about 100 ms per the documentation. In a phone agent every quarter second is felt, so Turbo is the natural choice for a live conversation, and full v4 for recorded segments.

Which ElevenLabs models support Hebrew

The practical trap: Flash v2.5 is the recommended, fastest model on ElevenLabs' Agents Platform, so it is the default for most people building an agent. In Hebrew it does not work. A vendor who did not explicitly choose v4 Turbo or v3 Conversational will give you a voice trying to read Hebrew in Latin letters. That is why “which voice model do you use” is the first question for a vendor.

Why Hebrew is hard for speech models

The difficulty splits into two sides: the voice that speaks and the understanding that listens. On the voice side the problems are pronunciation and phrasing; on the understanding side they are recognition. These are the most common, and each has a fix in the instructions or the settings.

• Unvocalized spelling: “ספר” can be “sefer” (book) or “sapar” (barber), and the model guesses from context. With first names and professional terms the guess fails more often.
• Gendered numbers: “shlosha tipulim” (three treatments) but “shalosh sha'ot” (three hours). Digits in text do not tell the model which form to pick.
• Times and dates: “14:30” must become “half past two”, and “12.11” must become “the twelfth of November”, by context.
• English inside Hebrew: “do you have WhatsApp?” needs a language switch mid-sentence without changing the voice.
• Acronyms: VAT (מע״מ), health fund (קופ״ח) and IDF (צה״ל) are read as words, not letter by letter.
• On the understanding side: Russian, Arabic and Amharic accents, and business terms the general model does not know.

How Azorly uses Eleven v4

Azorly's digital assistant is built on ElevenLabs' Agents Platform with an Israeli phone number. Our decisions were verified against the live API in October 2026.

• All 8 personas (Dana, Michal, Omer, Noa, Itai, Tomer, Tamar, Shira) are produced with the Eleven v4 model, with a per-assistant switch to v4 Turbo when latency matters more than timbre.
• The voices were picked from the ElevenLabs voice library among voices tagged as Hebrew, not English voices speaking Hebrew.
• Speech recognition runs in real time, and the business's keywords (treatment names, doctors' names, product names) are loaded from the characterization to improve recognition.
• Text normalization is on, so digits, times and dates are converted into Hebrew words before they are spoken.
• The language-model instructions are built from the business's characterization, including the assistant's gender, what must never be promised and when to hand over to a person.

7 tips for Hebrew prompts in a voice agent

1. Numbers in words, or normalization on. “The appointment is at half past four” sounds right; “the appointment is at 16:30” may sound like a code. Prices: “one hundred ninety-nine shekels”.
2. A fixed gender for the assistant. If she is female, every verb is feminine (speaks, records, transfers) in every instruction and every template. Mixing is the most common error.
3. Neutral address to the caller until the gender is clear: “may I have your full name?” instead of the masculine or feminine “tell me”.
4. Acronyms as spoken: “ma'am” or “value added tax”, “health fund” instead of the abbreviation. Gershayim in text do not guarantee a correct reading.
5. Short sentences and one question at a time. On the phone there is no scrolling back.
6. Israeli openers and closers: “Shalom, you have reached…”, “yom tov”, “Shabbat shalom” on Friday, “shavua tov” on Saturday night, “chag sameach”. They tell the caller they are at home.
7. A pronunciation dictionary for names and brands. ElevenLabs supports pronunciation dictionaries, and a doctor's or product's name pronounced wrong breaks trust faster than any other error.

The instructions to the language model must follow the same rules, because the text the model writes is what the voice reads out. A prompt written in English that asks for Hebrew answers usually produces translated Hebrew; a prompt written in Hebrew, with example sentences the way Israelis say them, produces speech that sounds local. At Azorly the instructions are built in Hebrew from the business's characterization.

For business owners: what it means in practice

You do not need to remember model names. You need to ask the vendor which voice model they use and insist on a Hebrew demo with your business's terms: the treatment name, the doctor's name, a time and a price. If the numbers sound right and the name is said the way you say it, the voice is fine. If something sounds like a tourist, the model probably does not support Hebrew.

At Azorly you can hear every persona before choosing, and call the demo number to test a full conversation. The voices are Eleven v4, and your business's terms go into the characterization in 10 minutes.