Block introduces Dessa, an open conversational voice AI model trained from scratch on human speech dynamics
Jack Dorsey's fintech company open-sourced a low-latency speech-to-speech foundation model designed to eliminate robotic stiffness by faithfully emulating natural disfluencies, pauses, and spontaneous vocal inflections.
Technology news
Jack Dorsey’s payment and technology conglomerate Block on Friday unveiled Dessa, a groundbreaking open-weights conversational audio foundation model trained from the ground up entirely on spontaneous human dialogue. Developed by Block’s applied artificial intelligence laboratory, Dessa represents a radical architectural departure from conventional synthetic voice pipelines, which typically rely on rigid text-to-speech engines that produce pristine, impeccably punctuated sentences that feel sterile and robotic in prolonged conversational interactions.
Unlike traditional cascaded voice systems that transcribe audio to text, run large language model inference, and synthesize audio outputs via separate models, Dessa operates as an integrated end-to-end speech-to-speech neural model. By training exclusively on unscripted recordings of natural conversations, the architecture natively captures the complex acoustic cues of human dialogue—including subtle thinking pauses, vocal disfluencies like “um” and “uh,” micro-hesitations, breathing patterns, and dynamic pitch modulations that signal irony, warmth, or mild uncertainty.
In demonstrations published by Block research lead Ragavan Thurairatnam, Dessa engaged in fluid, unscripted banter with human speakers, effortlessly handling overlapping speech and spontaneous interruptions without breaking latency. In direct comparative benchmarks against proprietary voice systems such as OpenAI’s GPT-4o voice mode, human listeners consistently rated Dessa as significantly more natural and approachable, specifically citing its willingness to stumble slightly, laugh off topic shifts, and avoid repetitive corporate enthusiasm.
Block has released the core weights and training documentation under a permissive open-source license, positioning Dessa as a versatile audio foundation layer for edge hardware, smart home hubs, and interactive commerce assistants. Hardware developers can run quantized variants locally on consumer Apple Silicon and high-end ARM chipsets, enabling private, zero-latency conversational voice interfaces that operate entirely on-device without streaming microphone audio to centralized cloud infrastructure.