Dessa: Block's Speech-Only AI Learns to Talk Like a Baby — and Runs on a MacBook

Introduction: The AI That Never Read a Word

Every voice assistant you've ever used — Siri, Alexa, ChatGPT's voice mode, Gemini Live — is secretly a text machine. Your speech gets transcribed to text, an LLM thinks about the text, and a separate model converts the answer back into speech. Somewhere in that pipeline, everything that makes your voice yours — the sarcasm, the sigh, the crack when you're upset — gets thrown away.

This week, Jack Dorsey's Block released Dessa, a research model that throws out the pipeline entirely. Dessa is what Block calls "the most authentic full-duplex speech model ever built," and it has a genuinely unusual pedigree: it was trained from scratch on nothing but raw human conversation audio. No text. No pretrained language model. No transcripts. It learned to talk the way babies do — by listening.

For anyone picking voice AI tools in 2026, Dessa is a preview of where the whole category is heading: audio-native models that hear emotion, take turns naturally, and run on hardware you already own.

What Is Dessa?

Dessa is a true speech-to-speech model — audio in, audio out, end to end. Block's small research lab (staffed by veterans of the original Dessa, a Toronto deep-learning startup, reporting directly to Dorsey) built it around a few capabilities that text-pipeline assistants physically cannot fake:

• Full-duplex conversation — it listens and talks at the same time, like a person. It knows when to cut in and when to backchannel ("yeah!", "got it!", "ooh") to show it's following you.
• Native emotional hearing — it hears and speaks sarcasm, laughter, hesitation, a frustrated sigh, a voice cracking with sadness. Not inferred from word choice; heard directly in the signal.
• Faster than real-time inference — no awkward wait between your sentence and its reply.
• It runs locally on a MacBook Pro M4 Max "or better" — no cloud round trip, no per-minute API bill.

Block's demos are live recordings with no steering: Dessa picks what to say, when to say it, and how — including its own voice and personality. In one published demo, it holds a conversation with GPT-Live-1, a third-party voice model, with both sides fully AI-generated.

Why Training on Speech Alone Is Brutal

The interesting engineering story is what Block didn't use. Modern LLMs learn from trillions of written words. Dessa, Block says, learned language from hundreds of times fewer words — because speech is a brutally inefficient and unstable teacher:

• Speech never repeats. Say the same sentence twice and the raw data comes out different every time; typed text is byte-identical.
• A spoken word is thousands of times bigger than the same word in text.
• The model had to learn more than language — accents, pauses, breathing, humor, inflection, and turn-taking all change what a sentence means. "I didn't say she stole my code" means something different depending on which word you stress.

And the compute story is just as unusual: the team figured out how to train Dessa on a single NVIDIA GB200 NVL72 — one rack-scale system, not a datacenter. That's the opposite of the frontier-lab scaling narrative, and proof that a very small team (six researchers plus interns) can still find unexplored territory by attacking a constraint everyone else routes around.

Dessa vs. the Text-Pipeline Voice Assistants

To be clear about what Dessa is not: it is a research model, not a product. Block's stated reasons for building it are curiosity, better human–AI understanding, and safety — not commerce. The Verge's hands-on noted it sometimes wanders off on tangents and once called a user "Ronaldo," and reactions split between technical admiration and discomfort with the companion-app framing. It is not about to replace your voice agent stack.

But the comparison is still instructive for tool buyers:

Today, if you're building anything real — customer support lines, meeting notetakers, content narration — you still want the mature stack: production voice agents like ElevenLabs for generation and cloning, AssemblyAI or Whisper-based transcription for understanding, Microsoft MAI Voice 2 and Speechify for natural narration. What Dessa signals is what that stack gets replaced by, eventually.

Local, Small-Team, Single-Chip: The Anti-Scale Story

Two parts of the Dessa announcement deserve more attention than they're getting. First, local execution: a full-duplex, emotionally aware conversation model running on a laptop means voice AI can be private, offline, and free at the point of use — the same shift we're watching in local coding models. Second, the efficiency claim: if a model can learn language from hundreds of times fewer words than an LLM needs, that challenges the assumption that intelligence requires web-scale text. Block explicitly frames the open research questions: How efficiently do speech-trained models learn facts? New languages? In-context learning?

The safety argument is the one Block leads with, and it's worth quoting the logic: AI keeps gaining power and independence while remaining nearly deaf to the emotion in our voices. "Imagine being in serious pain while a voice AI has no idea and just keeps going. That's the safety problem."

Why It Matters for the Voice AI Tools Market

Dessa won't show up in your tool stack this quarter. But three ripples will reach the market:

• The emotion gap becomes a differentiator. Expect voice tool vendors to start advertising paralinguistic understanding — detecting frustration, interrupting politely, matching your energy — because the research now shows it's buildable.
• Local voice models become credible. A MacBook-class conversation model erodes the assumption that voice AI must live in the cloud. Privacy-sensitive buyers will start asking where the audio is processed.
• Small labs keep out-innovating giants on contrarian bets. Six researchers at a payments company produced a world-first by training on speech only. The frontier isn't only at the trillion-dollar labs.

If you're evaluating voice tools today, the practical advice is unchanged: buy the mature stack. But when you renew in a year, ask each vendor one question Dessa made uncomfortable: does your assistant actually listen — or does it just read?

Frequently Asked Questions