# Block unveils Dessa, a speech model trained only on human speech

published: 2026-10-24T15:05:38.843Z
topic: infrastructure
source: https://the-decoder.com/blocks-new-ai-model-dessa-was-trained-entirely-on-human-speech-to-catch-emotions-text-based-models-miss/
url: https://aiagentnews.io/stories/block-unveils-dessa-a-speech-model-trained-only-on-human-speech

Block, the company behind Square and Cash App, has unveiled Dessa, an AI speech model that it says was trained entirely on human speech with no text data. Dessa processes audio directly instead of converting it to text first, and Block says this lets it pick up emotions, sarcasm, laughter and pauses in real time.

Block says Dessa is the first AI model trained entirely on human speech, with no text data, and that it was built from scratch. Instead of converting audio to text first, the model processes speech directly as audio. Block says this lets it detect emotions, sarcasm, laughter and pauses in real time.

Dessa supports full-duplex communication, so it can listen and talk at the same time. Block says it runs locally on a MacBook Pro M4 Max and was trained on a single NVIDIA GB200 NVL72. The company says it built the model out of curiosity, a wish for better human-AI communication, and concern that current voice AI ignores emotional cues.

Dessa is not publicly available yet. Block is first looking for AI security researchers to test it for unexpected risks, and people can sign up through a form. Earlier emotional voice systems include Hume AI's EVI, Kyutai's Moshi and OpenAI's Advanced Voice Mode.

## Why it matters if you build agents

Developers building voice agents cannot use Dessa yet, since access is limited to a sign-up form for security testers. If the claims hold, a speech model that works on audio directly could handle interruptions and emotional cues without a separate transcription step. Until there is public access and independent testing, the emotional-understanding claims rest on Block's own description.