The new model achieves a p95 time-to-first-audio latency of 150 milliseconds at 48 kHz, a speed intended to keep callers engaged without the robotic pauses that often trigger human intervention. Rather than licensing external technology, the Tel Aviv-based company trains its own models from scratch, allowing for rapid deployment of new languages within two weeks. Current support spans over 50 locales and dialects, verified by local language experts.

Deepdub is specifically addressing the failure points that cause enterprise voice agents to stall, such as misread dates or mangled surnames. In English, the engine employs advanced text normalization to translate raw data like '2024-12-31' or '$1,240' into natural speech. The most significant technical hurdle, however, remains Hebrew. Because the language is written without vowels, a single string of letters can carry multiple meanings depending on context. Phantom Z 3.4 resolves this by processing entire sentences to determine correct pronunciation rather than relying on word-by-word interpretation. This approach has earned the model top marks on the ivrit.ai leaderboard, with blind tests showing a 71 percent preference rate over the company's previous iteration.