{"id":603,"date":"2026-10-24T07:06:36","date_gmt":"2026-10-23T23:06:36","guid":{"rendered":"https:\/\/aidashxp.com\/block-dessa-speech-only-model\/"},"modified":"2026-10-24T07:07:43","modified_gmt":"2026-10-23T23:07:43","slug":"block-dessa-speech-only-model","status":"publish","type":"post","link":"https:\/\/aidashxp.com\/en\/block-dessa-speech-only-model\/","title":{"rendered":"Block Dessa \u6df1\u5ea6\u89e3\u6790\uff1a\u53ea\u9760\u8bed\u97f3\u8bad\u7ec3\u7684\u5168\u53cc\u5de5\u5bf9\u8bdd\u6a21\u578b\uff0cMacBook \u5c31\u80fd\u8dd1"},"content":{"rendered":"<p class=\"wp-block-paragraph\">Voice assistants always fall short: they only speak after you finish, cut off instantly when you interrupt, and sound completely flat. Block (founded by Jack Dorsey, parent company of Square) unveiled its answer on October 24 <strong>Dessa<\/strong>\u2014 A conversational model trained from scratch exclusively on human speech. No text, no transcription, no text-based pretraining\u2014not a single word was ever written down.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">This approach is fundamentally different from the path we\u2019re familiar with.<a href=\"https:\/\/chatgpt.com\" target=\"_blank\" rel=\"nofollow noopener\">ChatGPT<\/a>Gemini\u2019s voice capabilities are essentially a three-stage pipeline: \u201cspeech recognition \u2192 text-based large language model \u2192 speech synthesis,\u201d whereas Dessa has only one stage:<strong>Audio in, audio out<\/strong>Block\u2019s explanation is straightforward\u2014babies learn to speak without ever seeing written words, and they wanted to see whether a model could learn the same way.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">What exactly is Dessa?<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">The official definition is \u201ca true speech-to-speech model whose entire knowledge comes from human speech.\u201d Key features include:<\/p>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Full-duplex<\/strong>\u2014 It can listen while speaking, inserting interjections like \u201chmm\u201d or laughter mid-utterance, or even interrupting to correct itself\u2014just like humans\u2014rather than waiting for you to finish before processing.<\/li>\n<li><strong>Latency under 200 milliseconds<\/strong>\u2014 This figure approaches natural human conversational pauses; delays exceeding 500 milliseconds make users distinctly feel \u201cthe machine is waiting.\u201d<\/li>\n<li><strong>Conversational memory<\/strong>\u2014 Within a single session, it remembers prior exchanges and maintains contextual continuity, rather than resetting to zero with each turn.<\/li>\n<li><strong>Run locally<\/strong>\u2014 Officially stated to run locally on a MacBook Pro M4 Max or more powerful hardware\u2014no cloud connection required.<\/li>\n<li><strong>Native understanding of paralinguistic information<\/strong>\u2014 Tone, emotion, sarcasm, laughter, pauses\u2014elements that vanish entirely upon transcription into text\u2014are directly perceived by the model.<\/li>\n<\/ul>\n\n\n\n<h2 class=\"wp-block-heading\">Training data volume is just one-thousandth that of text-based models.<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">This is, to me, the most intriguing part. Text-based large models are typically trained on several trillion tokens, whereas Dessa\u2019s training volume, per official disclosure, is<strong>In the range of thousands of hours of speech<\/strong>Block\u2019s team claims that one hour of speech is roughly equivalent to a book\u2019s worth of text, and it would take ten billion hours to reach the trillion-word scale\u2014yet they\u2019ve slashed several zeros off that figure, betting on the premise that \u201cspeech itself carries higher information density.\u201d<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">In other words, Dessa is a<strong>reverse experiment<\/strong>: while everyone else is bolting speech capabilities onto text models, Block is attempting to prove that \u201clearning to speak only\u201d is viable. If successful, it implies low-resource languages, dialects, and colloquial expressions could be covered without massive textual corpora\u2014a prospect with considerable implications for Chinese dialect scenarios.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Demo performance: natural, but clearly immature<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">In several audio clips released officially, Dessa comes across as \u201can earnest but somewhat unpolished new friend\u201d: while discussing Nolan films, it abruptly shifts to Bryan Singer\u2019s Superman; when told it will be taken to Paris, it replies with something resembling \u201ca tearful thank-you\u201d; and it mistakenly identifies the speaker as someone named \u201cRonaldo\u201d at the start of a game.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">Another clip features Dessa conversing with OpenAI\u2019s GPT-Live-1 voice model; The Verge judges Dessa as sounding more natural. Yet, in my view, such officially curated samples are best taken with a grain of salt\u2014real-world experience requires hands-on use, and right now you<strong>cannot access it<\/strong>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Three critical limitations that must be clarified<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>First, it is not a product\u2014at present.<\/strong>Block explicitly states that Dessa is a research project; its weights are not open-sourced and access is granted only to a small cohort of AI safety researchers for vulnerability probing. Those hoping to download and run it locally should abandon the idea.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Second, it supports English only.<\/strong>The speech-only training approach has no known performance metrics for Chinese\u2014tones, erhua, and dialectal variations pose entirely new challenges for a model trained \u201con sound alone.\u201d No Chinese evaluation data exists yet.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\"><strong>Third, hardware requirements are substantial.<\/strong>A MacBook Pro with M4 Max chip is the minimum, implying unified memory of at least 48 GB. This mirrors the earlier situation with local large-model inference: technically feasible, yet out of reach for most users. For insights into such barriers, see <a href=\"https:\/\/aidashxp.com\/en\/meta-muse-spark-open-review\/\">Muse Spark Open benchmark<\/a> and <a href=\"https:\/\/aidashxp.com\/en\/recursive-flint-review\/\">Recursive Flint benchmark<\/a>\u2014both exemplify \u201cplug-and-play laptops\u201d whose entry thresholds remain prohibitive.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Where does Dessa sit within the voice AI landscape?<\/h2>\n\n\n\n<figure class=\"wp-block-table\"><table>\n<thead><tr><th>Product<\/th><th>Technical approach<\/th><th>Can it be used directly?<\/th><th>This Site\u2019s Rating<\/th><\/tr><\/thead>\n<tbody>\n<tr><td><strong>Block Dessa<\/strong><\/td><td>Pure speech training, full-duplex, runs locally<\/td><td>Researchers only, not open to the public<\/td><td>Not rated yet<\/td><\/tr>\n<tr><td><a href=\"https:\/\/aidashxp.com\/en\/gemini-live-avatar-face\/\">Gemini Live Avatar<\/a><\/td><td>Text LLM + real-time speech + video avatar<\/td><td>Available, supports 97 languages<\/td><td>8.6<\/td><\/tr>\n<tr><td><a href=\"https:\/\/aidashxp.com\/en\/chatgpt-voice-mobile-agent\/\">ChatGPT Voice Agent<\/a><\/td><td>Speech-to-text + Agent execution + text-to-speech<\/td><td>Available, cross-app operations<\/td><td>8.4<\/td><\/tr>\n<tr><td><a href=\"https:\/\/aidashxp.com\/en\/gpt-6-sol-luna-review\/\">GPT-6 Sol \/ Luna<\/a><\/td><td>Text flagship model; speech uses Live API<\/td><td>Available<\/td><td>8.9<\/td><\/tr>\n<\/tbody><\/table><\/figure>\n\n\n\n<p class=\"wp-block-paragraph\">The conclusion is clear:<strong>Dessa may lead on the single metric of \u201cnaturalness,\u201d but scores zero on usability<\/strong>Gemini Live already enables real-time video conversations in 97 languages,<a href=\"https:\/\/chatgpt.com\" target=\"_blank\" rel=\"nofollow noopener\">ChatGPT<\/a> and voice agents can check your email across apps\u2014these are things you can use today. Dessa\u2019s current value lies in proving the feasibility of a technical approach, not delivering a usable tool. On why voice AI has been slow to achieve the \u201calways-on\u201d moment, see this article<a href=\"https:\/\/chatgpt.com\" target=\"_blank\" rel=\"nofollow noopener\">ChatGPT<\/a> Industry interview analysis <a href=\"https:\/\/aidashxp.com\/en\/voice-ai-chatgpt-moment-not-yet\/\">Why this warrants close attention<\/a>.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Two reasons. First,<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">architectural differences drive experiential differences<strong>: Today\u2019s all voice assistants suffer \u201clag\u201d due to sequential ASR\u2192LLM\u2192TTS latency; Dessa\u2019s end-to-end single-model approach collapses the pipeline into one stage, theoretically approaching human responsiveness. Second,<\/strong>The \u201clag\u201d in all current voice assistants stems from the sequential delay across ASR\u2192LLM\u2192TTS; Dessa\u2019s single-model end-to-end approach compresses this pipeline into one stage, theoretically bringing it closer to human responsiveness. Second,<strong>Safety motivation<\/strong>Block states clearly that teaching models to speak from acoustics is a completely new domain\u2014so external researchers must first identify risks, including deception risks arising from outputs that \u201csound convincingly human but are actually machine-generated.\u201d<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">As for why Jack Dorsey built this, the team cites three reasons: \u201ccuriosity (i.e., fun), improving human\u2013AI communication, and safety.\u201d It sounds less like a commercial plan and more like Block\u2019s consistent ethos.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Frequently Asked Questions (FAQ)<\/h2>\n\n\n\n<h3 class=\"wp-block-heading\">Can Block Dessa be downloaded and used?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Not currently. Block positions it as a research project; model weights are not publicly released and access is restricted to a small group of AI safety researchers. The official website hosts only demo audio recordings, with no application portal.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">Does Dessa support Chinese?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">No Chinese-language capability data has been released officially; all demos are in English. Performance of pure speech-trained models on Mandarin tones and regional dialects remains unexplored, and expectations should be low in the near term.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">What is a full-duplex speech model?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Full-duplex means the model can simultaneously receive and generate audio: it listens while you speak and can interject with utterances like \u201cuh-huh\u201d or \u201cright\u201d at any moment\u2014you may even interrupt it mid-response. In contrast, half-duplex models operate strictly turn-by-turn: \u201cyou finish speaking, then it responds,\u201d allowing only one direction at a time.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">What hardware configuration is required to run Dessa?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">Officially, a MacBook Pro with M4 Max or better is recommended\u2014specifically devices with unified memory of 48 GB or higher. Even if weights are released later, most standard laptops will struggle to run it.<\/p>\n\n\n\n<h3 class=\"wp-block-heading\">How does Dessa differ from Gemini Live and<a href=\"https:\/\/chatgpt.com\" target=\"_blank\" rel=\"nofollow noopener\">ChatGPT<\/a> speech capabilities?<\/h3>\n\n\n\n<p class=\"wp-block-paragraph\">The key distinction lies in training methodology: Dessa uses speech-only input without text, whereas Gemini and <a href=\"https:\/\/chatgpt.com\" target=\"_blank\" rel=\"nofollow noopener\">ChatGPT<\/a> are large language models augmented with speech interfaces. In practice, Dessa feels more natural but supports only English and remains unavailable; the latter two offer broad capabilities, multilingual support, and immediate usability today.<\/p>\n\n\n\n<h2 class=\"wp-block-heading\">Summarize<\/h2>\n\n\n\n<p class=\"wp-block-paragraph\">Dessa is a valuable proof-of-concept\u2014not a product. It demonstrates that the path \u201clearning dialogue solely from speech\u201d is at least viable, while also establishing benchmarks for full-duplex operation, sub-200ms latency, and local execution. For users in China, the only actionable step now is observation\u2014awaiting weight release or adoption of this approach by a domestic vendor.<\/p>\n\n\n\n<p class=\"wp-block-paragraph\">To continuously track real-world usability and benchmark scores of new models, refer to our <a href=\"https:\/\/aidashxp.com\/en\/ai-models\/\">AI Model Library<\/a> and <a href=\"https:\/\/aidashxp.com\/en\/compare-tools\/\">Tool Comparison Engine<\/a>or continue reading:<a href=\"https:\/\/aidashxp.com\/en\/liquid-ai-lfm-2-5-24b-a2b\/\">Liquid AI LFM 2.5 Review<\/a> \u00b7 <a href=\"https:\/\/aidashxp.com\/en\/amazon-bee-ai-devices-frontier-eclipse\/\">Amazon Bee personal AI device<\/a> \u00b7 <a href=\"https:\/\/aidashxp.com\/en\/reviews\/\">All Reviews<\/a>.<\/p>","protected":false},"excerpt":{"rendered":"<p>\u8bed\u97f3\u52a9\u624b\u7528\u8d77\u6765\u603b\u5dee\u70b9\u610f\u601d\uff1a\u4f60\u8bf4\u5b8c\u5b83\u624d\u5f00\u53e3 [&hellip;]<\/p>\n","protected":false},"author":0,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"site-sidebar-layout":"default","site-content-layout":"","ast-site-content-layout":"default","site-content-style":"default","site-sidebar-style":"default","ast-global-header-display":"","ast-banner-title-visibility":"","ast-main-header-display":"","ast-hfb-above-header-display":"","ast-hfb-below-header-display":"","ast-hfb-mobile-header-display":"","site-post-title":"","ast-breadcrumbs-content":"","ast-featured-img":"","footer-sml-layout":"","ast-disable-related-posts":"","theme-transparent-header-meta":"","adv-header-id-meta":"","stick-header-meta":"","header-above-stick-meta":"","header-main-stick-meta":"","header-below-stick-meta":"","astra-migrate-meta-layouts":"default","ast-page-background-enabled":"default","ast-page-background-meta":{"desktop":{"background-color":"var(--ast-global-color-5)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"ast-content-background-meta":{"desktop":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"tablet":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""},"mobile":{"background-color":"var(--ast-global-color-4)","background-image":"","background-repeat":"repeat","background-position":"center center","background-size":"auto","background-attachment":"scroll","background-type":"","background-media":"","overlay-type":"","overlay-color":"","overlay-opacity":"","overlay-gradient":""}},"footnotes":""},"categories":[5],"tags":[],"class_list":["post-603","post","type-post","status-publish","format-standard","hentry","category-ai-audio"],"_links":{"self":[{"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/posts\/603","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/comments?post=603"}],"version-history":[{"count":1,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/posts\/603\/revisions"}],"predecessor-version":[{"id":607,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/posts\/603\/revisions\/607"}],"wp:attachment":[{"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/media?parent=603"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/categories?post=603"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/aidashxp.com\/en\/wp-json\/wp\/v2\/tags?post=603"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}