Financial technology firm Block has introduced a novel voice artificial intelligence model named Dessa. The project was developed by the company's in-house research team led by Ragavan Thurairatnam, which originated from Block's earlier acquisition of the startup Dessa. With this release, the enterprise backed by Jack Dorsey is pursuing an architectural path distinct from conventional voice interfaces. Instead of chaining separate speech-to-text and text-to-speech layers around a text foundation, Block built an architecture native to sound.
The core technical differentiator lies in the training pipeline. According to Block, Dessa was trained from scratch exclusively on audio recordings of human conversations, entirely bypassing written text datasets and pre-trained text language models. Conventional voice assistants typically transcribe user utterances to text, process that text through an underlying language model, and then generate speech. This sequential approach often introduces perceptible latency and strips away critical vocal cues such as intonation, laughter, and natural hesitations.
By processing audio directly, Dessa achieves genuine full-duplex interaction. The system can listen and speak simultaneously, enabling conversations that mirror human cadence. Users can fluidly interrupt the model mid-sentence, and Dessa can interject or pause without awkward latency buffers. Furthermore, the model naturally incorporates situational laughter and responsive backchanneling, including conversational affirmations like quick verbal acknowledgments.
The engineering team also focused on resource efficiency during development. Block stated that Dessa was trained on a single Nvidia rack featuring the GB200 NVL72 architecture. More significantly for end users, inference is designed to operate locally in real time on high-end consumer hardware, specifically demonstrated on a MacBook Pro powered by an M4 Max chip. This local execution removes round-trip network delays and eliminates dependency on remote cloud APIs.
At present, Block is not opening Dessa to the general public immediately. The system is restricted to a closed research preview managed through a security waitlist. Block's research division intends to evaluate potential misuse, vocal cloning risks, and edge-case behavior before planning wider commercial deployment. The release nevertheless highlights a broader shift toward end-to-end multimodal systems that treat spoken language as its own primary modality.