Israeli Startup Deepdub Launches Advanced AI Voice Model for Hebrew
Israeli startup Deepdub has released Phantom Z 3.4, a 3.4-billion-parameter AI voice model that outperforms major global tech giants in Hebrew text-to-speech benchmarks.
The Israeli startup Deepdub has unveiled its new AI voice model, Phantom Z 3.4, featuring 3.4 billion parameters and achieving top performance in Hebrew text-to-speech tasks.
End-to-End Technology
Unlike many competitors that rely on third-party technologies, Deepdub developed Phantom Z 3.4 entirely from scratch using proprietary datasets. Zivit Katz, Deepdub's Head of Marketing, explained that the company maintains full control over the pipeline, allowing them to add new languages to production environments within roughly two weeks.
"We do not rely on an external TTS model, but develop the technology entirely in-house from end to end," Katz noted.
Overcoming Hebrew Language Challenges
Hebrew presents unique obstacles for text-to-speech systems due to the lack of vowel markings in standard writing, where identical spelling can yield entirely different meanings depending on context. The model uses contextual analysis rather than isolated word reading to handle complex grammar, abbreviations, numbers, and proper nouns correctly.
In benchmark evaluations such as Coval and ivrit.ai's TTS Arena, Phantom Z 3.4 outperformed models from major global competitors including Google, OpenAI, and ElevenLabs. The proprietary model is currently available via the company's website and API, with commercial deployments already active in customer service automation, gaming, and advertising.
