Comparing Hebrew text-to-speech providers in 2025

Hebrew text-to-speech is one of those areas where the gap between English and other languages is still very real. I recently put together a snapshot comparing TTS quality across six providers, and the results ranged from impressively natural to completely unintelligible.

The test setup

I tested six providers: MiniMax (via Replicate voice cloning), Microsoft Edge TTS, Google Gemini, ElevenLabs, Chatterbox (via Replicate), and Resemble AI. For the voice cloning experiments, I used about one minute of English source audio per voice, created clones on Replicate, and then tried generating Hebrew speech with those cloned voices.

The winner: MiniMax

MiniMax using the T2A v2.6 Turbo model with voice clones and Hebrew boost enabled produced the most impressive results by far. The cloned voices sounded genuinely natural in Hebrew, which was a pleasant surprise given that the source audio was entirely in English.

The middle tier

Edge TTS (Microsoft's free offering with Avri and Hila voices), Gemini (via Google AI Studio with Puck and Zephyr voices), and ElevenLabs all produced good quality Hebrew. One critical gotcha with ElevenLabs: you must use the v3 model with an explicit language_code: "he" parameter. The multilingual v2 model produces completely unintelligible Hebrew output.

The disappointments

Chatterbox's voice cloning didn't carry through to Hebrew at all — the output sounded generic regardless of the cloned voice. Resemble AI needs nekudot (Hebrew diacritical marks) in the input text to produce intelligible output, which is a significant limitation since most modern Hebrew text doesn't include diacritics.

The repo

I've published the full comparison with audio samples, provider details, and links to useful Hebrew TTS resources on GitHub. If you're working with Hebrew speech synthesis, this should save you some trial and error.

Check it out: Hebrew-TTS-Providers on GitHub