{"id":1065,"date":"2026-10-24T18:22:49","date_gmt":"2026-10-24T18:22:49","guid":{"rendered":"https:\/\/perit.ai\/blogs\/?p=1065"},"modified":"2026-10-24T18:50:13","modified_gmt":"2026-10-24T18:50:13","slug":"block-dessa-speech-only-ai-conversation-model","status":"publish","type":"post","link":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/","title":{"rendered":"Dessa Learned to Talk Without Text. What Did It Learn to Understand?"},"content":{"rendered":"\n
\n

Block\u2019s speech-only AI puts the information between words at the center of the conversation. The next test is whether hearing that information changes what a model does.<\/p>\n

Picture a customer trying to explain a problem with an order. This is an illustrative call, the kind a voice agent might encounter tomorrow morning.<\/p>\n

\u201cI checked the delivery address, and\u2026\u201d<\/p>\n

\u201cYeah.\u201d<\/p>\n

\u201c\u2026it\u2019s still showing the old flat.\u201d<\/p>\n

Nothing unusual has happened. The listener has made a tiny sound to say: keep going, I\u2019m here. The customer continues without starting the sentence again. Neither person has formally handed over the conversation.<\/p>\n

Now imagine the same \u201cyeah\u201d arriving with more force, held a little longer, pressing into the sentence. The customer stops. The listener has something to say.<\/p>\n

A tidy transcript can make these exchanges look almost identical. The experience of being in them is different. In one, the acknowledgment supports the speaker. In the other, it asks for the floor. A system that treats every burst of speech as a new turn can mishandle both.<\/p>\n

This small event is a useful way into Dessa, Block\u2019s new conversational speech model<\/a>. Block says it trained Dessa from scratch using human conversational audio, without written training material or pretrained models. It takes audio in, produces audio out, and can listen while it speaks. The company presents its pauses, interjections, and expressive delivery as behavior learned from human interaction.<\/p>\n

The interesting promise reaches beyond a more convincing voice. Conversation carries information in when a sound arrives, what it overlaps, and how it is delivered. Retaining that information gives a model more evidence. The harder question is whether it learns to use that evidence appropriately.<\/p>\n

The conversation starts before the sentence ends<\/h2>\n

One of Block\u2019s featured recordings concerns a forgotten movie title. The publisher\u2019s embedded captions show the human asking for the second film\u2019s title, then Dessa beginning an answer before a later statement about having forgotten it. The exchange subsequently becomes confused about which sequel is under discussion.<\/p>\n

Those captions offer a revealing juxtaposition: anticipation and uncertainty in the same conversation. They do not establish how natural the overlap sounds. Their timestamps are publisher metadata, and subtracting them would not produce a measured response latency. The recordings have not been independently listened to or acoustically assessed for this article.<\/p>\n

Even with that boundary, the editorial question is clear. A model might join a conversation at an apparently appropriate moment while losing track of the thing being discussed. Timing and factual coherence deserve separate evidence.<\/p>\n

Most people already understand this from talking to other people. A quick answer can be wrong. A well-timed laugh can make an exchange easier without resolving a misunderstanding. Confidence in delivery can survive long after confidence in the underlying facts should have disappeared.<\/p>\n

Block\u2019s page invites attention to this layer of interaction. Its demonstrations include a volume-change request, a conversational surprise, life advice, and exchanges with another AI voice system. Block describes the recordings as live and unsteered, with captions added afterward. These are the publisher\u2019s stated conditions, rather than conditions independently verified here.<\/p>\n

The clips introduce a research direction. Establishing the boundaries of that direction requires a different kind of encounter: repeated conversations, disclosed conditions, and outcomes that can be checked.<\/p>\n

A transcript makes choices about what survives<\/h2>\n

Return to the customer and the old address. A transcription system might capture every word correctly. It could still leave the next component with a poor account of what happened.<\/p>\n

Did the listener\u2019s acknowledgment begin during \u201cdelivery address,\u201d during the pause after \u201cand,\u201d or after the entire sentence? Did the customer continue through it? Did the listener raise their voice to compete? Was the silence a completed thought or a search for the right word?<\/p>\n

An ordinary transcript often leaves those questions unanswered. Punctuation supplies an interpretation. A comma can stand where a breath, a hesitation, or a momentary interruption occurred. The printed sentence looks settled even when the speaker was still deciding how to finish it.<\/p>\n

That matters long before evaluation. Suppose a recording becomes a collection of polished, alternating utterances. Silence is trimmed. Overlap is removed. Repeated words disappear. Each speaker is assigned a neat block, followed by the next speaker\u2019s neat block.<\/p>\n

The resulting examples may be easier to read and manage. They may also contain fewer opportunities to learn what a listener does while another person is talking. A training pipeline has made a choice about the interaction available to the model.<\/p>\n

This is an inference about dataset design, not an explanation of Dessa\u2019s private training process. Block has not disclosed enough of that process to connect a particular preprocessing choice to a particular behavior.<\/p>\n

It also leaves room for many architectures. Some voice systems use speech recognition, a text model, and speech synthesis. Such systems can preserve timing or carry additional acoustic features. Native audio systems can work more directly with sound. The useful question for either approach is what evidence survives each transformation and reaches the component choosing the response.<\/p>\n

The presence of text on a screen tells us little about that boundary. A system can show a transcript while retaining audio. Another can present a voice interface while discarding cues that later matter.<\/p>\n

Different recordings. The same transcript.<\/div>
Flattening a conversation can make distinct interactions look identical.<\/div>
Recording A<\/div>
\u201cYeah\u201d arrives during speech<\/div>
Speaker<\/div>
<\/div><\/div><\/div>
Listener<\/div>
<\/div><\/div><\/div>
time \u2192<\/div> <\/div>
Recording B<\/div>
\u201cYeah\u201d arrives after speech<\/div>
Speaker<\/div>
<\/div><\/div><\/div>
Listener<\/div>
<\/div><\/div><\/div>
time \u2192<\/div> <\/div> <\/div>
\u2193<\/div>
BOTH FLATTEN TO<\/div>
Speaker: I changed the address.<\/div>
Listener: Yeah.<\/div><\/div>
These words alone cannot tell you which timing occurred.<\/div> <\/div>
Figure 1.<\/strong> A flat transcript is compatible with both timings. Time alignment can preserve the distinction. Bars are schematic; overlap alone does not establish intent.<\/figcaption> <\/figure>\n

The people who once made a voice convincing<\/h2>\n

There is a longer history behind Dessa\u2019s arrival.<\/p>\n

In May 2019, the original Toronto startup Dessa introduced RealTalk<\/a>, a text-to-speech imitation of Joe Rogan\u2019s voice. The team\u2019s account said the generated audio, including breaths and hesitation sounds, came from text input. Hashiam Kadhim and Rayhane Mama were among the engineers named in that work.<\/p>\n

The central provocation was the voice itself: software producing an identifiable person\u2019s speech. The same account discussed impersonation and misuse, and said the model and datasets would be withheld at that time.<\/p>\n

On 7 February 2020, Square announced its acquisition of Dessa<\/a>. The announcement emphasized the team\u2019s deepfake-detection work as well as its broader machine-learning experience.<\/p>\n

Six years later, Kadhim and Mama appear on Block\u2019s current Dessa team page. The research group is led by Ragavan Thurairatnam and, according to Block, reports directly to Jack Dorsey.<\/p>\n

The continuity of people makes the change in question compelling. The earlier public work showed how convincing a generated voice could become. The new project asks how a system might participate in the interaction surrounding that voice: arriving, yielding, acknowledging, and continuing.<\/p>\n

These are distinct projects. The historical connection does not establish shared architecture, training data, or technique. It does explain why this story has a longer runway than a company suddenly discovering that speech matters.<\/p>\n

There is a useful tension in that history. A voice can become persuasive before its claims become dependable. As systems gain more expressive control, the evidence behind their responses needs to become easier to inspect.<\/p>\n

\u201cOnly speech\u201d is a claim with several boundaries<\/h2>\n

Block\u2019s announcement combines three propositions that are easy to blur together: learning from speech, starting without pretrained models, and interacting in both directions at once.<\/p>\n

Each needs its own evidence. Speech-only training describes the supplied language experience. Training from scratch concerns initialization. Concurrent listening and speaking describes interaction. A model can satisfy one without satisfying the others.<\/p>\n

Speech can also be represented computationally in many ways. Discrete audio tokens would be compatible with learning without written supervision. Their presence would not imply a hidden text corpus. Conversely, the announcement gives no basis for assigning Dessa a particular codec, tokenizer, transformer design, or parameter count.<\/p>\n

A technical report would help establish how far \u201cfrom scratch\u201d extends across the complete system, including any audio representations and later training stages. For now, that remains a claim attributed to the team.<\/p>\n

The comparison with babies learning to talk is an accessible motivation. Children also see faces, manipulate objects, receive social feedback, and move through the world. Similarity in one input channel cannot establish similarity in the whole learning process.<\/p>\n

The availability boundary matters, too. In the official page markup checked for this article, the Download control still had no download URL, and a screen-reader note said launch links were forthcoming. The page offered an application for a small group of safety researchers. No linked technical report, public weights, API documentation, or reproducible evaluation harness was found in the examined launch material.<\/p>\n

Dessa is therefore best understood at this point as a research announcement with restricted access. That leaves an intriguing claim open for investigation. It also limits what an outside team can reproduce today.<\/p>\n

The past already contains textless dialogue<\/h2>\n

The novelty becomes sharper when earlier work enters the picture.<\/p>\n

In 2022, researchers introduced dGSLM, a generative spoken dialogue language model<\/a>. It trained on 2,000 hours of two-channel Fisher conversations without text labels and generated both channels, including overlapping speech and paralinguistic sounds.<\/p>\n

That precedent matters directly to this story. Textless dialogue learning and the modeling of conversational overlap were research problems before Dessa. Generated dialogue samples, however, do not automatically amount to a comparable live assistant.<\/p>\n

Moshi, introduced in 2024<\/a>, approached real-time dialogue with parallel audio streams and textual scaffolding. Its published design began with a text language-model backbone and predicted time-aligned text alongside speech. It shows why native audio systems should not all be described as simple transcription-and-synthesis pipelines.<\/p>\n

NVIDIA\u2019s PersonaPlex<\/a>, released in January 2026, provides another full-duplex precedent. Its Moshi-derived system listens and speaks concurrently, using an audio prompt for voice and a text prompt for role and context.<\/p>\n

These projects help separate the axes. Textless learning, live interaction, and application control can develop along different paths. Block\u2019s narrower claim concerns realistic live conversation learned from human speech without pretrained foundations. Assessing that claim will require clear component boundaries and comparable evidence of what \u201crealistic\u201d means.<\/p>\n

One rack tells only part of the story<\/h2>\n

Another launch detail is unusually easy to compress into the wrong headline. Block says Dessa trained on one NVIDIA GB200 NVL72.<\/p>\n

NVIDIA describes the NVL72<\/a> as a liquid-cooled rack containing 72 Blackwell GPUs and 36 Grace CPUs. \u201cOne\u201d identifies the rack. This is substantial training infrastructure.<\/p>\n

Its economic meaning depends on the workload and duration. The launch material examined here did not supply training time, effective GPU-hours, utilization, or total expenditure. A rack specification cannot answer those questions on its own.<\/p>\n

The local inference claim has a different boundary. Block names an M4 Max MacBook Pro or better and describes generation as faster than real time. That is a throughput claim. A system can produce audio faster than playback and still spend time deciding when to begin or stop.<\/p>\n

Numerical first-audio delay, interruption delay, memory requirements, and model precision were not supplied in the material examined. Nor does a newer chip generation automatically establish compatibility with a smaller laptop. Device fit and sustained performance will need measurements on the actual release.<\/p>\n

Block also says its model learned from dramatically fewer spoken words than typical LLM training. Without disclosed counts, a named baseline, and comparable capabilities, that remains an interesting claim rather than a demonstrated learning-efficiency multiplier. A corpus of speech and a corpus of written documents can differ in far more than word count.<\/p>\n

For teams considering deployment, the practical evidence would be a fixed model build on a named device, tested under a stated load, with memory use and latency distributions. That is where an appealing local-compute story becomes something an engineer can plan around.<\/p>\n

The voice has to change the decision<\/h2>\n

Now move the conversation from a movie discussion to an illustrative support call. The agent is explaining how to correct an address. Halfway through, the customer says, \u201cWait, that\u2019s the wrong order.\u201d<\/p>\n

A successful interaction requires several things to happen together. The system has to keep receiving audio while speaking. It has to recognize a correction, yield the turn, identify which order the customer means, and continue under the applicable rules. An appropriately apologetic tone helps the exchange, but the corrected order is the checkable outcome.<\/p>\n

This separates capabilities that a smooth demo can make feel like one. Detecting a cue is perception. Choosing a response involves the task and its constraints. Expressing that response is delivery.<\/p>\n

A model could recognize frustration and still repeat the instruction that caused it. It could stop speaking promptly, then resume the abandoned answer. It could preserve the correct amount while changing the wrong account. Each failure calls for a different diagnosis.<\/p>\n

A recent evaluation offers a useful way to examine part of this gap. VoxParity<\/a>, posted in September 2026, holds the spoken words fixed while varying acoustic evidence that should change a typed action. Its author reports that 11 of the 23 systems eligible for the transcript comparison passed its words-only null test.<\/p>\n

That is a protocol-specific result from one recent research report. The public development split uses synthetic audio, which limits conclusions about natural human conversation. Dessa was not established as evaluated in the work.<\/p>\n

The methodological connection is valuable: compare decisions when the words remain the same and only the relevant sound changes. Then examine whether the change in behavior is justified. Naturalness ratings alone cannot supply that answer.<\/p>\n

Does the sound change the action?<\/div>
Illustrative test rule: accept a reference only when every digit is audible.<\/div>
One reference transcript, held fixed<\/div>
\u201cThe reference is four seven.\u201d<\/div><\/div>
Audio A<\/div>
Final digit is clear<\/div>
four<\/span>seven<\/span><\/div>
\u2193<\/div>
EXPECTED ACTION<\/div>
Accept reference 47<\/div> <\/div>
Audio B<\/div>
Noise masks the final digit<\/div>
four<\/span>[masked]<\/span><\/div>
\u2193<\/div>
EXPECTED ACTION<\/div>
Ask for the last digit<\/div> <\/div> <\/div>
Text-only control:<\/strong> receives the same reference transcript for A and B. That input does not distinguish the recordings.<\/div>
Score the action change, not just whether the model mentions noise.<\/div> <\/div>
Figure 2.<\/strong> A proposed matched-pair test, not a VoxParity item or a Dessa result. The reference transcript is the underlying script, not an ASR output from the masked audio. The stated rule makes the action change justified.<\/figcaption> <\/figure>\n

Such tests also need cases where the system should continue normally. A model that escalates whenever a voice sounds unfamiliar can appear sensitive while creating another kind of failure. The goal is a warranted response to relevant evidence, with both missed cues and unnecessary interventions counted.<\/p>\n

The training example becomes a conversation<\/h2>\n

For a training-data company such as Perit, Dessa raises a practical question: what should a delivered example contain when the behavior being learned unfolds between speakers?<\/p>\n

An isolated utterance can be useful for pronunciation or recognition. A perfectly written transcript can be useful for lexical accuracy. Teaching or evaluating concurrent interaction may require a longer record: separate speaker channels, shared timing, silence, overlap, and enough context to understand why a brief sound occurred.<\/p>\n

Consider the \u201cyeah\u201d at the start of this article. Cut it away from the surrounding sentence and its role becomes harder to recover. Clean away the overlap and the listener\u2019s timing disappears. Compress the exchange into a summary and the acknowledgment may vanish altogether.<\/p>\n

The recording still needs curation. Breathing, laughter, microphone noise, and a second person in the room should not be casually conflated. Speaker diversity, recording conditions, usage rights, and session provenance affect what a dataset can support. Disagreements about an interruption\u2019s purpose should survive review where the evidence is ambiguous.<\/p>\n

None of this means speech-only pretraining requires human emotion labels. Raw audio can support learning without them. Annotation serves particular purposes: selecting examples, checking alignment, analyzing failures, designing evaluations, or supporting a later training stage. Those purposes should be stated.<\/p>\n

Perit\u2019s open conversation datasets<\/a> and live voice environments<\/a> make this connection concrete. Its published work includes channel-separated support speech and calls with trained speakers who interrupt or change their minds. The relevant perspective is preserving the interaction and scoring what the agent accomplishes. Perit is not identified here as a supplier to Dessa, and this article makes no claim about Block\u2019s undisclosed corpus.<\/p>\n

The lesson for a collection brief is simple to express and harder to execute: specify the behavior the data must make observable. If the problem is premature interruption, a delivery containing only polished alternating turns may omit the very event the team needs to study.<\/p>\n

A convincing conversation needs several scorecards<\/h2>\n

An evaluation should begin with ordinary human conversation and become more controlled where a specific failure needs isolation.<\/p>\n

First, test turn-taking with natural pauses that mean different things. Some speakers pause before continuing. Others finish without a long silence. Count premature cutoffs and unnecessary waits, then report their distributions across speakers and conditions.<\/p>\n

Next, introduce corrections while the system is speaking. Measure when it yields, but also inspect the response afterward. Did the new detail enter the answer? Did the abandoned request remain abandoned? Preserve the audio and the task trace so a reviewer can check both.<\/p>\n

Then test factual fidelity separately. Names, dates, amounts, and unfamiliar references deserve exact checks and opportunities to express uncertainty. Fluency should not compensate for losing a digit. Long conversations should return to earlier commitments and give the model opportunities to repair mistakes.<\/p>\n

Acoustic understanding needs its own matched comparisons. Change a relevant cue while keeping the words fixed. Include ordinary control cases. Use human recordings for final claims about real conversation, and label synthetic material where it helps isolate a variable.<\/p>\n

Finally, assess naturalness through blind listener comparisons, with consented recordings, and report it separately from correctness. An expressive voice may be preferred even when its answers are less accurate. Both findings matter to a product team.<\/p>\n

Application control is another open question. Before testing role adherence or tool actions, an evaluator needs a documented interface that supports them. A demo titled \u201cRefund\u201d does not establish a payments integration. A conversation between two AI systems does not establish a benchmark victory.<\/p>\n

These are proposed tests, not Dessa results. Their purpose is to make progress legible: better timing, more faithful answers, appropriate actions, or a more pleasant interaction. Each is worth knowing about on its own terms.<\/p>\n

Back to the little word<\/h2>\n

The customer is still trying to explain the old address. The listener says \u201cyeah,\u201d and the customer keeps going.<\/p>\n

For a person, that can be an almost invisible success. For a learning system, it is an event distributed across two streams of sound. The meaning depends on what came before, what is happening at the same moment, and what happens next.<\/p>\n

Dessa makes that event central to its research story. Block says a model learned to participate by listening to human conversation, without the written foundation familiar from text-trained assistants. The public recordings invite a closer look. A fuller release will make it possible to ask stronger questions about how that behavior was learned and how reliably it holds.<\/p>\n

The questions reach beyond Dessa. What did preprocessing preserve? Which speakers and conditions appeared in the data? When the model heard a correction, did it change course? When the recording sounded reassuring, did the task actually end well?<\/p>\n

The evidence will live in complete interactions: the pause left intact, the overlapping acknowledgment, the correction heard in time, and the outcome a person can verify. That is where a voice that keeps a conversation moving begins to earn trust with the work inside it.<\/p>\n

\n

Source note: Research and release controls checked on 24 October 2026. This article is based on public-source review, including Block\u2019s page markup and publisher-supplied captions. No Dessa model was installed or tested; demonstration audio was not independently listened to or assessed. Capability descriptions remain attributed to Block. Opening and support-call scenarios are illustrative. The proposed evaluations are not collected results.<\/em><\/p>\n<\/div>\n\n","protected":false},"excerpt":{"rendered":"

Block\u2019s speech-only AI puts the information between words at the center of the conversation. The next test is whether hearing that information changes what a model does.<\/p>\n","protected":false},"author":2,"featured_media":1062,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[22,19],"tags":[26],"class_list":["post-1065","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-artificial-intelligence","category-research","tag-ai-engineering"],"yoast_head":"\nBlock Dessa: Speech-Only AI and What Conversation Teaches It<\/title>\n<meta name=\"description\" content=\"Block says Dessa learned conversation from speech alone. Explore its origins, release status, and what voice AI must prove beyond natural delivery.\" \/>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/\" \/>\n<meta property=\"og:locale\" content=\"en_US\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"Block Dessa: Speech-Only AI and What Conversation Teaches It\" \/>\n<meta property=\"og:description\" content=\"Block says Dessa learned conversation from speech alone. Explore its origins, release status, and what voice AI must prove beyond natural delivery.\" \/>\n<meta property=\"og:url\" content=\"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/\" \/>\n<meta property=\"og:site_name\" content=\"Perit Blog\" \/>\n<meta property=\"article:published_time\" content=\"2026-10-24T18:22:49+00:00\" \/>\n<meta property=\"article:modified_time\" content=\"2026-10-24T18:50:13+00:00\" \/>\n<meta property=\"og:image\" content=\"https:\/\/perit.ai\/blogs\/wp-content\/uploads\/2026\/10\/block-dessa-speech-only-ai-thumbnail.webp\" \/>\n\t<meta property=\"og:image:width\" content=\"1600\" \/>\n\t<meta property=\"og:image:height\" content=\"840\" \/>\n\t<meta property=\"og:image:type\" content=\"image\/webp\" \/>\n<meta name=\"author\" content=\"Pratyaksh@Perit\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:creator\" content=\"@baldwin_IVth\" \/>\n<meta name=\"twitter:label1\" content=\"Written by\" \/>\n\t<meta name=\"twitter:data1\" content=\"Pratyaksh@Perit\" \/>\n\t<meta name=\"twitter:label2\" content=\"Est. reading time\" \/>\n\t<meta name=\"twitter:data2\" content=\"17 minutes\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\\\/\\\/schema.org\",\"@graph\":[{\"@type\":\"Article\",\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/#article\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/\"},\"author\":{\"name\":\"Pratyaksh@Perit\",\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/#\\\/schema\\\/person\\\/6b951f8ca802e00473088655d8a2a1bc\"},\"headline\":\"Dessa Learned to Talk Without Text. What Did It Learn to Understand?\",\"datePublished\":\"2026-10-24T18:22:49+00:00\",\"dateModified\":\"2026-10-24T18:50:13+00:00\",\"mainEntityOfPage\":{\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/\"},\"wordCount\":3320,\"commentCount\":0,\"image\":{\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/block-dessa-speech-only-ai-thumbnail.webp\",\"keywords\":[\"AI Engineering\"],\"articleSection\":[\"Artificial Intelligence\",\"Research\"],\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"CommentAction\",\"name\":\"Comment\",\"target\":[\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/#respond\"]}]},{\"@type\":\"WebPage\",\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/\",\"url\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/\",\"name\":\"Block Dessa: Speech-Only AI and What Conversation Teaches It\",\"isPartOf\":{\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/#website\"},\"primaryImageOfPage\":{\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/#primaryimage\"},\"image\":{\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/#primaryimage\"},\"thumbnailUrl\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/block-dessa-speech-only-ai-thumbnail.webp\",\"datePublished\":\"2026-10-24T18:22:49+00:00\",\"dateModified\":\"2026-10-24T18:50:13+00:00\",\"author\":{\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/#\\\/schema\\\/person\\\/6b951f8ca802e00473088655d8a2a1bc\"},\"description\":\"Block says Dessa learned conversation from speech alone. Explore its origins, release status, and what voice AI must prove beyond natural delivery.\",\"breadcrumb\":{\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/#breadcrumb\"},\"inLanguage\":\"en-US\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/\"]}]},{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/#primaryimage\",\"url\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/block-dessa-speech-only-ai-thumbnail.webp\",\"contentUrl\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/wp-content\\\/uploads\\\/2026\\\/10\\\/block-dessa-speech-only-ai-thumbnail.webp\",\"width\":1600,\"height\":840,\"caption\":\"Dessa editorial cover: It learned to talk. Did it learn to listen? Two overlapping audio traces highlight conversation beyond words.\"},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/block-dessa-speech-only-ai-conversation-model\\\/#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"Home\",\"item\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"Dessa Learned to Talk Without Text. What Did It Learn to Understand?\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/#website\",\"url\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/\",\"name\":\"Perit Blog\",\"description\":\"Frontier training data, collected and graded by people who do the work.\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/?s={search_term_string}\"},\"query-input\":{\"@type\":\"PropertyValueSpecification\",\"valueRequired\":true,\"valueName\":\"search_term_string\"}}],\"inLanguage\":\"en-US\"},{\"@type\":\"Person\",\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/#\\\/schema\\\/person\\\/6b951f8ca802e00473088655d8a2a1bc\",\"name\":\"Pratyaksh@Perit\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"en-US\",\"@id\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/wp-content\\\/uploads\\\/2026\\\/09\\\/pfp-150x150.webp\",\"url\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/wp-content\\\/uploads\\\/2026\\\/09\\\/pfp-150x150.webp\",\"contentUrl\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/wp-content\\\/uploads\\\/2026\\\/09\\\/pfp-150x150.webp\",\"caption\":\"Pratyaksh@Perit\"},\"sameAs\":[\"http:\\\/\\\/perit.ai\\\/blogs\",\"https:\\\/\\\/x.com\\\/baldwin_IVth\"],\"url\":\"https:\\\/\\\/perit.ai\\\/blogs\\\/author\\\/pratyakshpatel\\\/\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"Block Dessa: Speech-Only AI and What Conversation Teaches It","description":"Block says Dessa learned conversation from speech alone. Explore its origins, release status, and what voice AI must prove beyond natural delivery.","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/","og_locale":"en_US","og_type":"article","og_title":"Block Dessa: Speech-Only AI and What Conversation Teaches It","og_description":"Block says Dessa learned conversation from speech alone. Explore its origins, release status, and what voice AI must prove beyond natural delivery.","og_url":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/","og_site_name":"Perit Blog","article_published_time":"2026-10-24T18:22:49+00:00","article_modified_time":"2026-10-24T18:50:13+00:00","og_image":[{"width":1600,"height":840,"url":"https:\/\/perit.ai\/blogs\/wp-content\/uploads\/2026\/10\/block-dessa-speech-only-ai-thumbnail.webp","type":"image\/webp"}],"author":"Pratyaksh@Perit","twitter_card":"summary_large_image","twitter_creator":"@baldwin_IVth","twitter_misc":{"Written by":"Pratyaksh@Perit","Est. reading time":"17 minutes"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"Article","@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/#article","isPartOf":{"@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/"},"author":{"name":"Pratyaksh@Perit","@id":"https:\/\/perit.ai\/blogs\/#\/schema\/person\/6b951f8ca802e00473088655d8a2a1bc"},"headline":"Dessa Learned to Talk Without Text. What Did It Learn to Understand?","datePublished":"2026-10-24T18:22:49+00:00","dateModified":"2026-10-24T18:50:13+00:00","mainEntityOfPage":{"@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/"},"wordCount":3320,"commentCount":0,"image":{"@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/#primaryimage"},"thumbnailUrl":"https:\/\/perit.ai\/blogs\/wp-content\/uploads\/2026\/10\/block-dessa-speech-only-ai-thumbnail.webp","keywords":["AI Engineering"],"articleSection":["Artificial Intelligence","Research"],"inLanguage":"en-US","potentialAction":[{"@type":"CommentAction","name":"Comment","target":["https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/#respond"]}]},{"@type":"WebPage","@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/","url":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/","name":"Block Dessa: Speech-Only AI and What Conversation Teaches It","isPartOf":{"@id":"https:\/\/perit.ai\/blogs\/#website"},"primaryImageOfPage":{"@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/#primaryimage"},"image":{"@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/#primaryimage"},"thumbnailUrl":"https:\/\/perit.ai\/blogs\/wp-content\/uploads\/2026\/10\/block-dessa-speech-only-ai-thumbnail.webp","datePublished":"2026-10-24T18:22:49+00:00","dateModified":"2026-10-24T18:50:13+00:00","author":{"@id":"https:\/\/perit.ai\/blogs\/#\/schema\/person\/6b951f8ca802e00473088655d8a2a1bc"},"description":"Block says Dessa learned conversation from speech alone. Explore its origins, release status, and what voice AI must prove beyond natural delivery.","breadcrumb":{"@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/#breadcrumb"},"inLanguage":"en-US","potentialAction":[{"@type":"ReadAction","target":["https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/"]}]},{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/#primaryimage","url":"https:\/\/perit.ai\/blogs\/wp-content\/uploads\/2026\/10\/block-dessa-speech-only-ai-thumbnail.webp","contentUrl":"https:\/\/perit.ai\/blogs\/wp-content\/uploads\/2026\/10\/block-dessa-speech-only-ai-thumbnail.webp","width":1600,"height":840,"caption":"Dessa editorial cover: It learned to talk. Did it learn to listen? Two overlapping audio traces highlight conversation beyond words."},{"@type":"BreadcrumbList","@id":"https:\/\/perit.ai\/blogs\/block-dessa-speech-only-ai-conversation-model\/#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https:\/\/perit.ai\/blogs\/"},{"@type":"ListItem","position":2,"name":"Dessa Learned to Talk Without Text. What Did It Learn to Understand?"}]},{"@type":"WebSite","@id":"https:\/\/perit.ai\/blogs\/#website","url":"https:\/\/perit.ai\/blogs\/","name":"Perit Blog","description":"Frontier training data, collected and graded by people who do the work.","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/perit.ai\/blogs\/?s={search_term_string}"},"query-input":{"@type":"PropertyValueSpecification","valueRequired":true,"valueName":"search_term_string"}}],"inLanguage":"en-US"},{"@type":"Person","@id":"https:\/\/perit.ai\/blogs\/#\/schema\/person\/6b951f8ca802e00473088655d8a2a1bc","name":"Pratyaksh@Perit","image":{"@type":"ImageObject","inLanguage":"en-US","@id":"https:\/\/perit.ai\/blogs\/wp-content\/uploads\/2026\/09\/pfp-150x150.webp","url":"https:\/\/perit.ai\/blogs\/wp-content\/uploads\/2026\/09\/pfp-150x150.webp","contentUrl":"https:\/\/perit.ai\/blogs\/wp-content\/uploads\/2026\/09\/pfp-150x150.webp","caption":"Pratyaksh@Perit"},"sameAs":["http:\/\/perit.ai\/blogs","https:\/\/x.com\/baldwin_IVth"],"url":"https:\/\/perit.ai\/blogs\/author\/pratyakshpatel\/"}]}},"_links":{"self":[{"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/posts\/1065","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/comments?post=1065"}],"version-history":[{"count":4,"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/posts\/1065\/revisions"}],"predecessor-version":[{"id":1077,"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/posts\/1065\/revisions\/1077"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/media\/1062"}],"wp:attachment":[{"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/media?parent=1065"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/categories?post=1065"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/perit.ai\/blogs\/wp-json\/wp\/v2\/tags?post=1065"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}