Hebrew TTS Snapshot

A snapshot of Hebrew text-to-speech capabilities as of 22 March 2025, comparing voice quality across multiple TTS providers — including voice cloning experiments via Replicate.

Key Findings

• MiniMax — Voice cloning (Replicate) · Best · Most impressive results — cloned voices sounded natural in Hebrew

• Edge TTS — Stock voices (Avri, Hila) · Good · Microsoft's free TTS; tested at 100% and 70% speed

• Gemini — Stock voices (Puck, Zephyr) · Good · Via Google AI Studio, Gemini 2.5 Flash Preview TTS

• ElevenLabs — Stock voices (v3 model) · Good · Requires language_code: "he" — multilingual v2 is unintelligible

• Chatterbox — Voice cloning (Replicate) · Poor · Output was generic; cloning didn't carry through to Hebrew

• Resemble AI — Stock + voice cloning · Poor · Needs nekudot (diacritics) for intelligible output

Voice Cloning Method

Voice clones for MiniMax and Chatterbox were generated using Replicate:

1. ~1 minute of English source audio per voice (see voice-sources/english/)

2. Voice clone IDs created on Replicate from those samples

3. Hebrew text generated using the cloned voice IDs with Hebrew language parameter

4. Hebrew reference audio also tested (see voice-sources/hebrew/)

MiniMax used the T2A v2.6 Turbo model with voice clones and Hebrew boost enabled.

Repository Structure

├── samples/ # Generated TTS audio output │ ├── chatterbox/ # Chatterbox multilingual voice cloning (via Replicate) │ │ ├── run1/ # English reference audio, unvowelised input │ │ └── run2/ # Hebrew reference audio │ ├── edge-tts/ # Microsoft Edge TTS │ │ ├── avri-100pc/ # Avri voice, normal speed │ │ ├── avri-70pc/ # Avri voice, 70% speed │ │ ├── hila-100pc/ # Hila voice, normal speed │ │ └── hila-70pc/ # Hila voice, 70% speed │ ├── elevenlabs/ # ElevenLabs v3 model (Rachel, Adam, Bella) │ ├── gemini/ # Google Gemini 2.5 Flash Preview TTS (Puck, Zephyr) │ ├── minimax/ # MiniMax T2A v2.6 Turbo voice clones (Corn, Herman) │ └── resemble/ # Resemble AI │ ├── stock/ # Avigail (Hebrew preset voice) │ └── voice-clone/ # Chatterbox multilingual clone (Herman) ├── voice-sources/ # Input audio used for voice cloning │ ├── english/ # ~1 min English samples (corn, daniel, herman) │ └── hebrew/ # Hebrew reference samples │ ├── corn/ │ ├── daniel/ │ └── herman/ ├── texts/ # Hebrew text prompts │ ├── source/ # Texts used for TTS generation (+ PDF versions) │ └── target/ # Additional test texts (cooking, music, weather) └── resources.md # Links to related tools and services

Provider Details

MiniMax (Best Results)

• Model: T2A v2.6 Turbo

• Method: Voice clones created on Replicate, Hebrew boost enabled

• Voices tested: Corn, Herman

Edge TTS

• Voices: Avri (male), Hila (female)

• Speed variants: 100% (normal) and 70% (slowed)

• Texts: Family, Jerusalem, sample text, travel, and more

ElevenLabs

• Model: eleven_v3 with language_code: "he"

• Voices: Rachel (F), Adam (M), Bella (F)

• Note: Multilingual v2 (eleven_multilingual_v2) produces unintelligible Hebrew. v3 with explicit Hebrew language code is required.

Google Gemini

• Model: Gemini 2.5 Flash Preview TTS

• Interface: Google AI Studio

• Voices: Puck, Zephyr

Chatterbox (via Resemble/Replicate)

• Engine: Chatterbox Multilingual

• Method: Voice cloning via Replicate API with language: "he"

• Result: Output sounded generic — voice cloning did not carry through to Hebrew

• Issue in Run 2: Dicta ONNX model not loaded; output was still gibberish

Resemble AI

• Stock voice: Avigail (Hebrew preset) — usable but limited

• Custom clones: Unintelligible without nekudot (diacritics) in input text

Resources

• Phonikud TTS — Hebrew TTS with automatic diacritization

• Add Diacritics in Hebrew — Diacritics preprocessor

• LightBlue TTS Hebrew — Open source Hebrew TTS

• Ivrit.ai — Israeli speech technology

• Replicate TTS Collection — TTS models on Replicate