Fintech firm Block has introduced Dessa, a native speech AI model trained entirely on human voice data without relying on text representations. By processing raw audio directly rather than converting speech to text, the model captures real-time emotional cues, sarcasm, laughter, and pauses. It supports full-duplex communication, allowing simultaneous listening and speaking.
Dessa was trained on a single NVIDIA GB200 NVL72 system and operates locally on Apple silicon. Block has not released the model publicly, opting first to recruit external AI security researchers to conduct risk and safety testing.
Why it matters
Represents a shift toward audio-native architectures that bypass text-based intermediate steps for lower latency and better emotional context.
Demonstrates non-traditional AI players developing proprietary, specialized hardware-optimized models locally.
Offers new approaches for real-time customer service and voice interface applications.
Source: the-decoder.com



