IA · 24 October 2026 · 4 min read
Trained purely on speech without written text: Block unveils Dessa
In brief: Block, the technology company led by Jack Dorsey, has introduced Dessa, a frontier voice AI model trained from scratch exclusively on human speech recordings. Unlike conventional architectures that convert spoken audio into text strings before processing, Dessa extracts reasoning and emotional nuances directly from raw sound. Its full-duplex framework allows natural interruptions and instant cadence, while local execution on high-end laptops guarantees low latency and private data processing.
by Team Mocchi's
For years, conversational voice interfaces have relied on a fragmented, three-tier architecture: speech-to-text engines transcribe user audio, large language models generate written tokens, and text-to-speech synthesizers convert that text back into sound. While this approach has made vast web-scale knowledge accessible, it inherently strips oral communication of its richest layers: tone, pacing, hesitation, and natural timing. Jack Dorsey's Block has set out to overhaul this pipeline with Dessa, an artificial intelligence model trained from scratch strictly on human speech.
Beyond transcription: an audio-native foundation
The driving philosophy behind Dessa mirrors human language acquisition. As the Block research group led by Ragavan Thurairatnam pointed out, children learn to converse fluidly long before reading or writing their first word. The engineering goal was to determine whether a neural network could develop coherent world understanding and conversational reflexes solely from acoustic waveforms. According to reporting from TechCrunch, Dessa acquires factual associations and logic directly from conversational audio data, dispensing entirely with intermediate text layers.
This architecture bypasses the synthetic emotional tags and manual prompts used by traditional voice generators to fake expressiveness. With Dessa, vocal nuances such as laughter, sighs, pitch variation, and pauses emerge natively from the learned acoustic signal, sidestepping the robotic monotony that characterizes conventional virtual assistants.
Full-duplex interaction and conversational nuance
Architecturally, Dessa is a full-duplex system, meaning it continuously listens and speaks along a concurrent audio stream. Users can interject, correct instructions mid-sentence, or talk over the assistant without forcing it through an artificial turn-taking loop.
Public demonstrations shared by the team highlight diverse scenarios, ranging from casual guessing games to unscripted debates about cinema. As highlighted by The Verge, the model exhibits an unusual level of conversational naturalness, though it occasionally displays the uncanny eccentricities of early generative systems, such as unexpected tangents or momentary confusion. Jack Dorsey and his researchers have emphasized that flattening speech into text poses safety hazards: text-only layers remain deaf to critical emotional distress signals like panic or physical pain, making native acoustic understanding a fundamental safety prerequisite.
Local inference on consumer hardware
Beyond its training philosophy, Dessa's operational efficiency stands out. Despite handling rich audio streams in real time, Block designed the model to run entirely on local, consumer-grade hardware. The researchers confirmed that inference requires only a single Nvidia GPU or an Apple Silicon laptop such as a MacBook Pro with an M4 Max chip.
Running models edge-locally eliminates cloud round-trip network latency, the primary technical bottleneck that has previously hindered real-time spoken interactions. Furthermore, edge inference offers strict data confidentiality, ensuring that private voice conversations never traverse third-party servers. Access to Dessa is currently restricted to select AI safety researchers while Block evaluates integration across its product ecosystem.
Mocchi's take
The move toward audio-native, end-to-end speech models like Dessa points to the definitive future of human-machine interaction, moving past the awkward delay of traditional transcription-synthesis pipelines. For companies building interactive agents, automated field services, or hands-free industrial interfaces, on-device execution delivers unmatched real-time responsiveness alongside robust data sovereignty at reasonable hardware costs. For software engineers, the central challenge will involve pairing this spontaneous acoustic fluency with deterministic business logic, ensuring that high emotional nuance does not compromise operational reliability.