Jack Dorsey’s financial technology company Block has introduced Dessa, a new artificial intelligence voice model designed to hold natural conversations using speech alone. Announced on October 23, the model can listen and respond simultaneously while producing human-like expressions, including laughter, pauses and changes in emotional tone.
According to Block’s official announcement, Dessa was trained entirely on recorded human conversations without relying on written text or pretrained language models. The company said the model was developed using a single NVIDIA GB200 NVL72 computing rack and can operate locally on consumer hardware.
Dessa Processes Conversations Without Text Conversion
Unlike conventional voice assistants that convert speech into text before generating responses, Dessa processes audio directly. The model uses a full-duplex system, allowing it to listen while speaking, interrupt conversations and provide verbal acknowledgments without waiting for someone to finish talking.
This approach is similar to the real-time conversational capabilities introduced in OpenAI’s GPT-Live models. However, Block said Dessa develops its conversational knowledge directly from speech recordings rather than relying on a separate large language model.
The company argued that traditional speech-to-text systems lose important conversational information, including emotional cues, timing and vocal expressions. By training directly on spoken interactions, Block aims to preserve these characteristics during AI conversations.
In demonstrations published by its researchers, Dessa displayed different speaking styles, emotional expressions and conversational rhythms. One example featured the model sounding tired while discussing work, while another showed it responding emotionally during a conversation.
Other demonstrations included a guessing game between a Block employee and the voice assistant. The recordings illustrated how Dessa adjusts its speaking patterns to different conversational situations, although independent evaluations of its performance remain limited.
Block Targets Local Processing and Safety Testing
Block said Dessa can run locally on devices equipped with consumer-grade hardware, including a MacBook Pro with an M4 Max processor or better. This would allow compatible devices to process voice conversations without requiring constant access to remote servers.
However, Dessa is currently available only as a research preview, with no public download or general access. The company has invited a limited group of AI safety researchers to evaluate the model and identify potential risks associated with realistic synthetic speech.
Block has not disclosed the model’s parameter count, independent benchmark results or a public release date. Its claims about conversational performance and local processing have therefore not been independently verified through widespread testing.
Block already operates voice-enabled services through Square, including tools that help businesses manage customer orders. Although the company has not announced plans to integrate Dessa into its commercial products, its existing payment and business software operations provide potential applications for conversational voice technology.
Featured image credits: Wikimedia
For more stories like it, click the +Follow button at the top of this page to follow us.