Breaking01 / AI
AI model outputs first speech audio in 390 ms on NVIDIA GB10
A single model takes text, images and speech as inputs and uses a reference voice to synthesise speech.
Researchers reported in an arXiv paper on 20 Nov 2026 that DRVEXO-4B-Unified took 390 ms to produce its first chunk of synthesised speech on NVIDIA GB10 hardware.
The arXiv abstract describes a model that accepts text, images and speech. It uses a reference voice to synthesise speech.
The researchers said they planned to release the model weights and code under Apache License 2.0 and retain applicable third-party terms.
In the researchers’ published results for IFEval, an instruction-following benchmark, the model scored 82.26% under strict scoring at the prompt level. The abstract cited a published score of 81.2% for Qwen3-4B with reasoning mode turned off.
The researchers did not give a release date for the weights and code in the abstract.