[{"data":1,"prerenderedAt":14},["ShallowReactive",2],{"artikel-en-berita-block-dessa-model-suara-23-oktober-2026":3},{"slug":4,"lang":5,"title":6,"description":7,"date":8,"updated":9,"topic":10,"product":9,"cover":11,"readMin":12,"html":13},"berita-block-dessa-model-suara-23-oktober-2026","en","Dessa, Block's voice model that learned to talk without text","Block launched Dessa, a full-duplex voice model trained only on human speech that runs locally on a MacBook.","2026-10-25",null,"berita","\u002Fimages\u002Fartikel\u002Fberita-block-dessa-model-suara-23-oktober-2026.webp",3,"\u003Cp>Jack Dorsey's Block announced Dessa on October 23, 2026. It is a voice model trained from scratch using recordings of human conversation only. No text, no pretrained models. Block says it is the first AI you can have a realistic live conversation with that was trained entirely from speech.\u003C\u002Fp>\n\u003Cp>Unlike most voice AI, which converts your speech to text first, Dessa processes audio straight to audio. That is why it can pick up tone, sarcasm, laughter and a frustrated sigh, all of which disappear in a transcript.\u003C\u002Fp>\n\u003Ch2>1. What Dessa can do\u003C\u002Fh2>\n\u003Cp>Dessa is a full-duplex model, which means it listens while it speaks, the way people do. It learned from real conversations when to cut in and when to backchannel, for example saying "yeah" or "got it".\u003C\u002Fp>\n\u003Cp>In the demo recordings on Block's site, Dessa laughs mid-sentence when something is funny, speaks louder when asked, and completes a movie title the speaker forgot. All of that emerged on its own, without being trained for it specifically.\u003C\u002Fp>\n\u003Ch2>2. Trained without text, on a single rack\u003C\u002Fh2>\n\u003Cp>Block says Dessa learned language from roughly hundreds of times fewer words than a typical LLM. A single spoken word carries thousands of times more bytes than the same word in text, which is what made this hard.\u003C\u002Fp>\n\u003Cp>The striking part: Block says it trained Dessa on just one NVIDIA GB200 NVL72 rack. The model then runs locally on a MacBook Pro M4 Max or better, and inference is faster than real time.\u003C\u002Fp>\n\u003Ch2>3. Its knowledge comes from conversation alone\u003C\u002Fh2>\n\u003Cp>Voice models are usually paired with a large LLM for knowledge and retrieval. Dessa is not. TechCrunch reports that Block confirmed Dessa's knowledge comes entirely from the conversational data it was trained on.\u003C\u002Fp>\n\u003Cp>That means it is not a replacement for assistants like Siri or Google Assistant. It is a research model showing what is possible when a machine learns language through listening alone, the way babies do.\u003C\u002Fp>\n\u003Ch2>4. Not public yet, but safety researchers are invited\u003C\u002Fh2>\n\u003Cp>The public cannot try Dessa yet. Block has opened a waitlist and is inviting a small group of AI safety researchers to test it and look for risks its own team has not anticipated.\u003C\u002Fp>\n\u003Cp>The Implicator also pushed back on the claims: there is no published Dessa evaluation or external benchmark, and the showcased recordings were selected by Block itself. Independent testing will have to confirm how good the model actually is.\u003C\u002Fp>\n\u003Ch2>5. Context: the voice AI race is heating up\u003C\u002Fh2>\n\u003Cp>Dessa arrives after OpenAI released its GPT-Live voice models mid-year, and Kyutai's Moshi already supports full-duplex audio. In a demo on Block's site, Dessa talks with GPT-Live-1, a third-party model, without awkward pauses.\u003C\u002Fp>\n\u003Cp>The differentiator is the training setup: all speech, from the start. If that claim holds up under independent testing, it could change how we judge voice AI, from how accurate the answers are to how human the conversation feels.\u003C\u002Fp>\n\u003Ch2>What it means for you\u003C\u002Fh2>\n\u003Cp>Voice AI today usually feels slow and stiff because it turns your speech into text first. Dessa points to another path, and it runs on an ordinary laptop without any cloud round trip.\u003C\u002Fp>\n\u003Col>\n\u003Cli>\u003Cstrong>Expect more natural voice assistants.\u003C\u002Fstrong> If the speech-only approach works, voice AI will read your emotion and pacing, not just your words.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>You will not need a data center GPU.\u003C\u002Fstrong> Block says Dessa runs on a MacBook M4 Max. Local voice AI is becoming realistic for everyday devices.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>Do not trust official demos blindly.\u003C\u002Fstrong> Dessa's recordings were picked by Block and there is no independent benchmark yet. Wait for outside evaluation.\u003C\u002Fli>\n\u003Cli>\u003Cstrong>There is no product yet.\u003C\u002Fstrong> Dessa is a research preview with no price or release date. For today's needs, try tools already listed on \u003Ca href=\"\u002Fen\u002Fai-tools\">AI Tools\u003C\u002Fa>.\u003C\u002Fli>\n\u003C\u002Fol>\n\u003Ch2>Sources\u003C\u002Fh2>\n\u003Cul>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fblock.xyz\u002Fdessa\" target=\"_blank\" rel=\"noopener\">Dessa, the Conversation Model trained only on speech\u003C\u002Fa> (Block)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Ftechcrunch.com\u002F2026\u002F10\u002F24\u002Fblocks-new-voice-model-sounds-eerily-human\u002F\" target=\"_blank\" rel=\"noopener\">Block's new voice model sounds eerily human\u003C\u002Fa> (TechCrunch)\u003C\u002Fli>\n\u003Cli>\u003Ca href=\"https:\u002F\u002Fwww.implicator.ai\u002Fblock-dessa-speech-only-voice-model\u002F\" target=\"_blank\" rel=\"noopener\">Awkwardly Natural: Dessa Sounds Like Someone You Just Met\u003C\u002Fa> (The Implicator)\u003C\u002Fli>\n\u003C\u002Ful>\n",1793543302057]