Block’s voice AI learned to talk without ever reading a word

Dessa's official recording player shows separate human and AI speech tracks and labeled conversation moments

Block’s voice AI Dessa was trained entirely on human speech, without text or a previously trained model, according to the company. The research system can listen and speak at the same time, but early access is limited to a small group of AI safety researchers.

Block, which owns Square and Cash App, has published company demonstrations and says Dessa can run locally on a MacBook Pro with an M4 Max chip or better.

Dessa keeps the sound behind the words

Dessa processes speech directly. Block says this lets the model respond to vocal cues such as laughter, pauses and sarcasm that a plain transcript can lose.

Listening while speaking, often called full-duplex conversation, allows overlapping speech and brief acknowledgments instead of requiring each speaker to finish a turn. Block is seeking outside safety testing before wider access to this speech-only approach.

Block’s demonstration page presents separate speech tracks. The recording is a company example, not an SEW product test. Image: Block

A natural voice still needs a working business connection

For businesses, a convincing conversation is one component of an automated service. Completing an order also requires the correct menu, payment process and connection to the system that sends work to staff.

Square’s existing AI-powered voice ordering documentation makes that dependency concrete. Customers receive a payment link by text, and their order reaches the point-of-sale or kitchen system only after checkout is completed.

That service uses a Square business phone number and an ordering profile. Its documentation lists Square Plus, Square Premium and Square Messages Plus subscribers; Block’s Dessa announcement gives no timetable for connecting the research model to this ordering workflow.

A studio microphone, headphones and audio mixer under colored lighting
Illustrative audio-recording equipment, not Block’s Dessa setup. Image: Los Muertos Crew / Pexels

The separation is visible in Mistral’s Voxtral 3 voice-agent architecture, which connects speech with a reasoning model and business tools. Those components provide the records and actions that a conversation needs; Dessa’s announcement focuses on learning and conducting the conversation itself.

Source: Block

The post Block’s voice AI learned to talk without ever reading a word appeared first on Search Engine Watch.

Source:: searchenginewatch.com