Block's Dessa model learns to chat from raw speech, skipping text entirely
Jack Dorsey's Block unveiled Dessa, an AI model trained exclusively on human conversation that operates as a speech-to-speech system, capable of running locally on high-end MacBooks.
Last verified:
Speech-to-speech training from conversation alone
Block, the financial services and cryptocurrency company founded by Jack Dorsey, unveiled Dessa, a conversational AI model trained exclusively on human speech rather than text transcriptions. According to The Verge, Dessa operates as a true speech-to-speech system, meaning it learns and generates language entirely from audio patterns without intermediate text conversion—a departure from the text-first pipeline most commercial voice AI systems use today.
The model’s training methodology mirrors how human infants acquire language: through immersion in spoken dialogue alone. As Block’s Dessa team explained to The Verge, “Babies learn to talk without ever seeing a word written down. We wanted to see if a model could too. It can. It’s just very hard.” This approach aims to preserve prosody, timing, and emotional nuance that text normalization would otherwise discard.
Preserved timing and emotional recognition in conversations
The Dessa team demonstrated the model’s conversational capabilities through recorded examples, including a guessing game played over voice and discussions ranging from Christopher Nolan’s filmography to travel anecdotes. According to The Verge’s coverage, some exchanges display natural rapport, though occasional quirks emerge—in one instance, Dessa initially misidentified a speaker as “Ronaldo.” In another example, Block described Dessa responding to a Paris trip anecdote “with what sounds like teary gratitude,” suggesting the model captures emotional tone in its responses.
A central design principle, outlined by the Dessa team to The Verge, addresses what the researchers frame as a safety gap: modern voice AI systems “flatten what we say into text and can’t keep up with human timing,” and crucially, they remain “nearly deaf to the emotion in our voices.” Dessa’s speech-native architecture aims to detect and respond to vocal emotion—relevant in scenarios where a user is in pain or distress and requires emotionally aware AI support.
Local deployment and early research access
Dessa runs locally on high-end consumer hardware; The Verge reports the model can execute on a MacBook Pro M4 Max or better, eliminating dependence on cloud infrastructure and addressing latency and privacy concerns tied to centralized inference. Block cited three motivations for developing Dessa: “curiosity (aka fun), better understanding between humans and AI, and safety.”
The company is restricting initial access to a small cohort of AI safety researchers, suggesting a cautious rollout focused on stress-testing the model’s emotional reasoning and failure modes before broader deployment.
Why This Matters
Voice-first AI development has historically treated speech as a format to be transcribed and normalized rather than preserved. If Dessa’s approach of training directly on audio proves practical at scale, it could reshape how conversational AI handles prosody-dependent tasks—customer support interactions where tone matters, therapeutic applications, and accessibility interfaces where naturalness reduces cognitive load. For teams building voice AI products, Dessa signals that text-as-intermediary may no longer be the only viable path.
The open question is whether Dessa’s emotional detection generalizes beyond the curated demo examples The Verge observed. Independent evaluation by the early-access researchers will determine whether the model genuinely recognizes distress cues or merely mimics conversational naturalness in scripted scenarios. If the former, the safety case for speech-native training strengthens; if the latter, Block’s foundational premise—that text normalization destroys safety-relevant information—requires recalibration.
Frequently Asked Questions
What makes Dessa different from other voice AI models?
Dessa is trained exclusively on human speech rather than text transcriptions, preserving emotional tone and conversational timing. It operates as a true speech-to-speech system without converting audio to text intermediaries.
Can I run Dessa on my own computer?
Yes, according to Block, Dessa can run locally on a MacBook Pro M4 Max or better, enabling offline inference without cloud dependency.
Who can access Dessa right now?
Block is initially granting early access to a small group of AI safety researchers; broader availability has not been announced.