Hebrew AI voiceover: Eleven v3 vs v4 vs MiniMax Speech

The short answer

To turn Hebrew text into a voiceover, pick one of the five models whose official language list includes Hebrew, paste the text, choose a voice and create. The result is an audio file saved to your library. The price follows the number of characters, and this is what 1,000 characters cost right now:

According to the maker of Eleven v3, 5,000 characters are about 5 minutes of speech, so 1,000 characters are about one minute of voiceover.

The five models side by side

Three other Eleven models, Eleven Multilingual v2, Eleven Turbo v2.5 and Eleven Flash v2.5, do not include Hebrew on their official language lists. The Eleven name alone does not promise Hebrew, so for a Hebrew voiceover pick only one of the five models in the table.

Which model fits what

Eleven v3: narration with emotion

A speech model built for human, highly expressive speech. Tags in English inside square brackets, such as [whispers] or [sighs], change how a line is read mid sentence. According to its maker it is meant for recorded narration rather than real-time conversation, and it takes a little more trial and error with wording than earlier models.

Eleven v4: the new generation

According to its maker, Eleven v4 is built to read tone, pacing, emotion and character from the text, and keeps the same voice and character across a long project. Its tags can also add sounds, such as a laugh or light rain in the background. On an independent text to speech leaderboard from October 2026 it ranks 2nd of 96. That ranking is general and does not test Hebrew in particular.

Eleven v4 Turbo: the same generation made fast

The same technology as Eleven v4, and according to its maker it starts speaking within about 150 milliseconds. On the same independent leaderboard it ranks 1st of 96. Here you get a finished file rather than a live conversation, so speed matters less, and its price is ₪0.191.88 crEleven v4 Turbo, 1,000 chars per 1,000 characters, against ₪0.383.75 crEleven v4, 1,000 chars in Eleven v4.

MiniMax Speech 2.8 HD and Turbo: full control of the voice

Both models officially support Hebrew through the "Language boost" field. You pick an emotion, such as happy, sad or angry, and tune speed from 0.5 to 2, pitch and volume. According to the maker you can add sounds inside the text such as (laughs), (sighs) or (breath), and an exact pause with a mark such as <#0.5#>, where the number is the pause length in seconds. One request takes up to 10,000 characters, so a long text goes in at once.

The Turbo version gives the same voices and the same options: ₪0.282.82 crMiniMax Speech 2.8 Turbo, 1,000 chars per 1,000 characters in MiniMax Speech 2.8 Turbo, against ₪0.474.69 crMiniMax Speech 2.8 HD, 1,000 chars in MiniMax Speech 2.8 HD.

So which is best in Hebrew

We have not yet run a structured listening comparison of the five models on the same Hebrew text, so there is no single winner here. What is known: all five officially support Hebrew.

The best way to choose is to read the same short paragraph in two or three models and decide by ear. A 300 character paragraph costs ₪0.141.41 crEleven v3, 300 chars in Eleven v3, ₪0.111.13 crEleven v4, 300 chars in Eleven v4 and ₪0.090.85 crMiniMax Speech 2.8 Turbo, 300 chars in MiniMax Speech 2.8 Turbo right now.

How to prepare Hebrew text for reading aloud

• Tell the model it is Hebrew. In the Eleven models, write he in the "Language code" field, or leave it empty and the model detects the language. In the MiniMax Speech models, pick Hebrew in the "Language boost" field
• Niqqud is optional. Write plain Hebrew text. We have not tested systematically how each model reads Hebrew without niqqud, so listen for words that can be read two ways and spell them with extra vowel letters when needed
• Numbers and abbreviations. The Eleven models have a "Text normalization" setting that turns numbers, dates and abbreviations into words before reading. When a number is read wrong, write it out in words
• Long text in parts. Each request takes up to 5,000 characters in the Eleven models and up to 10,000 in the MiniMax Speech models. Split longer text by paragraphs
• Emotion through tags. The tags are written in English inside the Hebrew text. The makers document them, but we have not yet tested them systematically in Hebrew text, so start with one tag and check
• Stability. In the Eleven models a high stability value gives an even, calm reading, and a low value gives more emotion and variety, with a less predictable result

Choosing a voice

In every model you choose from preset voices, each with its own sound and character. Cloning your own voice is not available here. We have not tested every voice in Hebrew, so it is worth trying two or three voices on one sentence before reading a long text. A 100 character sentence costs ₪0.040.38 crEleven v4, 100 chars in Eleven v4 right now, and it saves a long reading in a voice that does not fit.

Hebrew narration for a video

No video model here officially supports Hebrew speech, so the safe way to a video with Hebrew narration is to create the video and the voiceover separately and join them in an editing app. You can also upload the voiceover to Wan 2.7 as a driving audio file, and the model times lip movement and action to it, but there is no official confirmation that lip sync is accurate in Hebrew, and we have not tested it.

The full picture of Hebrew in video, images and music is in the guide Hebrew in AI models. With five models that officially support Hebrew and an exact price before every reading, Hebrew voiceover is now the easy part of the project.

Questions and answers