Block has released Dessa, a voice AI it says was trained from scratch on conversational audio alone, as a research preview. According to the company, training used neither text data nor pretrained models. Dessa takes in speech directly and responds with speech, and it supports full-duplex conversation, meaning it can speak while it listens. Processing audio directly and learning language without text are separate technical features, however. What stands out about Dessa is not only its natural spoken dialogue, but its attempt to acquire language and knowledge from conversational audio without relying on written text.
Learning vocal inflection and conversational timing
In the demos Block published, Dessa laughs while the other person is still speaking, raises its voice when asked to speak louder, and fills in the words for someone trying to remember a movie title. Each example shows a system that reacts as the conversation unfolds, unlike conventional voice dialogue, which generates a reply only after the other person has finished speaking.
Full-duplex means the model can process the other party's speech and produce its own at the same time. It doesn't need to wait for the speaker to finish completely before offering a backchannel response or cutting in. Direct audio input and output, by contrast, is a design choice about how speech is received and produced. Combining the two means the dialogue model handles not just the content of words but also vocal inflection, laughter, and timing.
Plain text cannot adequately capture things like a sigh, a pause, or a sarcastic tone. Block says Dessa processes this kind of audio-specific information and decides for itself when to offer backchannel responses or interject. The team also adopted a policy of using "neither text nor pretrained models" in training. Rather than adding voice capabilities to a language model trained on text, the attempt is to have the model acquire language from human conversational audio itself.
Block says the conversation clips it published are recordings of real exchanges between a person and Dessa. The model itself decided what to say, when to say it, and how to voice it, with no human operating behind the scenes; captions were added after recording. Still, these are demos chosen by the developer, and the same quality is not guaranteed in ordinary conversation. Even if the model can produce a tearful-sounding reply or a surprised gasp, that alone does not prove the AI has emotions or correctly understands the other person's feelings.
Trained on 72 GPUs, runs inference on a MacBook Pro
According to Block, training Dessa used one rack of NVIDIA's GB200 NVL72. Per NVIDIA's official specifications, the GB200 NVL72 is a liquid-cooled system linking 72 Blackwell GPUs and 36 Grace CPUs with a high-speed interconnect. The phrase "one rack" may suggest a small training setup, but the compute involved is far beyond that of an ordinary PC.
Inference, on the other hand, meaning actually conversing with the trained Dessa, can reportedly run locally on a MacBook Pro with an M4 Max or better. Training a model with heavy compute and running the finished model are separate stages with different hardware needs. The statement that it runs on a MacBook Pro describes only the inference environment; it does not mean training requires little compute.
Block also says Dessa's inference can run faster than real time. But speed of generating audio is a different metric from the delay between the other person finishing speaking and the first response beginning. The published research summary does not list specific response-start latency, the measurement conditions, or required memory. Because the parameter count and quantization conditions are also unknown, it is impossible to tell how comfortably the model runs on supported Macs.
On training data scale, Block claims Dessa acquired language from a few hundredths of the number of words typical large language models (LLMs) are trained on. It acknowledges, though, that a direct comparison is difficult because LLMs also learn programming languages such as C++. The comparison models, the absolute number of training words, and evaluation results based on equivalent capability have not been disclosed. Training duration, total GPU hours, and cost are also unpublished, so the smaller word count cannot simply be tied to lower training cost.
Versus the earlier Moshi, the difference is how it learns
Full-duplex voice dialogue itself was not first achieved by Dessa. The French AI research lab Kyutai announced the full-duplex voice AI Moshi in 2024. The original preprint by Alexandre Défossez and colleagues (arXiv:2410.00037v2) describes a design that processes the user's and the AI's audio as two parallel streams, without sharply separating listening time from speaking time. Moshi also handles audio input and output directly, so the ability to listen and speak at once is not unique to Dessa.
The major difference between the two lies in the starting point for learning language. Moshi was built on Helium, a 7-billion-parameter text language model. By Block's account, Dessa uses no pretrained text language model and learns language from human conversational audio alone. In other words, both handle audio directly in full duplex, but their philosophies differ on how language ability is acquired.
Based on Block's research overview of Dessa and the original Moshi paper dated October 2, 2024 (pages 1–3), the main differences can be summarized as follows.
| Comparison | Dessa (Block) | Moshi (Kyutai) |
|---|---|---|
| Full-duplex voice dialogue | Described as supported | Supported |
| Direct audio input/output | Described as supported | Supported |
| Training with text | Described as not used | Built on a text language model |
This comparison rests on design information published by each developer and paper's authors, not on performance evaluation under identical conditions. It therefore cannot establish which model is better in response speed, factual accuracy, or conversational ability.
Why Moshi uses text matters for judging Dessa's approach as well. Moshi's team adopted a method it calls "Inner Monologue," which first predicts text aligned in time with the speech, and says this improved factual accuracy and linguistic quality. That shows a model using text can still process audio directly, and suggests that a text-free design is not necessarily better in dialogue quality. For Dessa, a key question is how well it can maintain accuracy of content and linguistic ability without such text support.
Knowledge and safety remain open questions beyond natural speech
Block says it is offering early access to a small number of AI safety researchers to uncover risks it did not anticipate during development. The company also lists improving mutual understanding between humans and AI, and safety, as goals. As one concrete concern, it mentions that an AI might fail to recognize a voice expressing distress. However, no quantitative results have been published on how accurately Dessa can detect crisis situations or whether its safety improves on earlier voice AI.
The ability to respond naturally to a speaker's tone and the ability to return appropriate content for the situation need to be evaluated separately. For example, a model might answer in a surprised tone to someone's voice without correctly judging what there is to be surprised about. And fluent replies in a short conversation do not show that facts and context can be maintained accurately over a long one. Whether Dessa's audio-only approach can support practical dialogue depends on verifying such abilities.
Details of the audio data used for training are another important thing to check. Block's research overview does not disclose the total hours of conversational audio used, its sources, or its licensing status. The makeup of the data, including what kinds of conversations are included and in what proportion, is also unknown. To evaluate language ability and handling of emotional speech, it will be necessary to understand the characteristics of the training data and to confirm performance in conversations beyond the developer's chosen demos.
For now, Dessa remains a research preview, and a page for general users is being prepared. Japanese-language support, plans for an API, and pricing have not been announced, and it is unclear whether the model weights will be released. The technical significance of Dessa lies in trying to acquire natural conversational ability from audio alone, without a text language model as its base. If accurate knowledge in long conversations and appropriate responses to voices expressing distress are demonstrated, the potential of learning language from speech alone can be assessed beyond the natural speaking style shown in the demos.