Phantom Z 4.0: from 250 ms to 50 ms
Our new text-to-speech model for voice agents reaches 1.2% word error rate at a 58 ms median time-to-first-audio, putting it on the latency–accuracy Pareto frontier of Coval’s independent benchmark.
Same benchmark, same measurement, one month. On Coval’s independent text-to-speech leaderboard, the median time it takes Phantom Z to start speaking fell from roughly 250 ms to about 50 ms, network round trip included.
Today we’re introducing Phantom Z 4.0, the model behind that drop. It sits on Coval’s word-error-rate-vs-latency Pareto frontier with the lowest word error rate among the fastest TTS models: 1.2% WER at a 58 ms median time-to-first-audio. On tail latency it ranks second, at 71 ms p90 and 86 ms p95.
Phantom Z 4.0 is available now in early access in the Deepdub Playground, in English and Hebrew, with the rest of our supported languages to follow soon.
Fast, without trading away accuracy
Speed only matters if the words come out right. Coval checks intelligibility by transcribing each model’s audio with a fixed speech recognition model and scoring it against the input text. Plotted against latency in the chart at the top of this post, Phantom Z 4.0 lands on the Pareto frontier: no model on the leaderboard is both faster and more accurate.
Rebuilt from the ground up
The name continues the Phantom family, but almost everything underneath is new. Over the past two months, the team at Deepdub took the model from first principles to production: modeling, training, finetuning and the inference engine. The goal was to bring what Deepdub has learned in seven years of Hollywood dubbing to voice agents.
Built from scratch, drawing on recent advances in language and vision modeling.
Kernel fusion, quantization, dedicated Triton kernels and custom scheduling.
Engine-only TTFA holds at 25 ms for a single request or a burst of 16 concurrent streams.
What it delivers
Full-bandwidth output, ready for production.
Performance that follows the meaning of the text, not just its words.
Numbers, dates, abbreviations and spelled-out codes are read the way a person would say them.
Precise control over delivery, suited to production work.
How we got there
Two things made the difference.
- Data. Seven years of professional, multilingual performance data from real dubbing productions.
- Human feedback. Phantom Z 4.0 was RL-tuned on thousands of hours of human-rated audio.
“Designing the model and the inference engine together, rather than optimizing one after the other, is what let us push latency down without giving up quality.”
Chen, Chief Scientist
What’s next
Over the next few weeks we’re shipping:
- More kernel optimizations.
- A 24 kHz telephony variant.
- Distilled flash models.
Together, these should take roughly another 10 ms off time-to-first-audio and raise throughput further.
Try Phantom Z 4.0
Pick a use case, edit the text and press Play. Every request below runs on Phantom Z 4.0.
Acknowledgments. Thank you to Ohad Zohar, Tamir Cohen, Nir Krakowski and Yuval Dror, and to everyone at Deepdub who made Phantom Z 4.0 possible.
Benchmark figures come from Coval’s independent Voice AI Benchmarks, where the model is listed as “Phantom Z 3.4 conversational.” Leaderboard values are a snapshot from October 2026; the live leaderboard updates continuously.
