San Francisco: Block, the technology company led by Jack Dorsey, has unveiled Dessa, a voice artificial intelligence model designed to learn and communicate through human speech. The model aims to make conversations with AI sound more natural by reproducing features such as pauses, changes in volume, laughter and interruptions that occur during everyday exchanges.
Block says Dessa was trained on audio alone, without text or pretrained models. Unlike conventional voice systems that convert spoken words into text before generating a response, the model is designed to learn directly from human speech and retain elements of how people communicate, rather than focusing exclusively on the words being spoken.
The company has released audio demonstrations featuring conversations that range from natural banter to awkward misunderstandings. Some exchanges sound expressive and spontaneous, while others reveal mistakes that highlight the challenges of teaching AI systems to understand people through speech alone.
Dessa is described as a speech-to-speech model that determines what to say, when to speak and how to deliver its responses, including the voice and personality it adopts. Block says the project is intended to explore how AI can better understand human communication and what can be learned by training models directly on spoken interactions.
Dessa aims to capture the rhythm of human conversation
Human conversations involve much more than words. People change their tone, pause to think, speak over one another, laugh in response to unexpected comments and adjust their volume depending on the situation.
Block argues that voice systems that rely on converting speech into text can lose some of this information. Although a transcript can preserve the words spoken, it may not fully represent the timing, emotion or delivery that shaped the original exchange.
Dessa is designed to preserve these characteristics by learning from audio directly. According to the company, the model can respond with pauses, variations in volume and laughter that may begin before the other speaker has finished talking.
The system uses what Block describes as a full-duplex approach. This allows it to handle overlapping interactions, rather than requiring each speaker to finish an entire turn before the other responds.
The approach is intended to make exchanges feel less like a sequence of separate questions and answers and more like an ordinary conversation. In real-life interactions, listeners frequently react before a sentence is complete, interrupt to clarify a point or use laughter to signal their response. Capturing these behaviours is one of the challenges Dessa is designed to investigate.
Block says the model independently determines not only the content of a response but also its timing and vocal delivery. The demonstrations are intended to show how these decisions affect the flow of a conversation.
Film discussions and awkward misunderstandings
The audio demonstrations released by the team provide examples of both the model’s conversational abilities and its limitations.
In one recording shared by Dessa lead Ragavan Thurairatnam, the model initially mistakes its conversation partner for someone named Ronaldo. Despite the mistaken identity, the exchange continues in a largely natural manner.
The incident illustrates a distinction between sounding conversational and correctly understanding the context of a conversation. An AI system may respond with appropriate timing and an expressive voice while still misunderstanding who it is speaking to.
Another demonstration begins with a discussion of films by Christopher Nolan before moving unexpectedly towards Bryan Singer’s Superman films. The shift highlights how conversational systems can move between topics in ways that may appear spontaneous but can also seem disjointed.
A separate recording features a person telling Dessa about a surprise trip to Paris. The model responds in a way that the team describes as sounding tearfully grateful.
These examples show why Block is placing emphasis on audio delivery as well as conversational content. A response can communicate a particular emotion through tone, timing and volume, even when the words alone do not fully explain the effect.
However, natural-sounding speech does not necessarily establish that a model understands emotions in the same way humans do. The recordings demonstrate the behaviour of the system during specific exchanges, while broader questions about its capabilities remain subjects for research.
A conversation with OpenAI’s GPT-Live-1 voice model
Dessa’s website also features a conversation between the model and OpenAI’s GPT-Live-1 voice model.
The demonstration places two conversational AI systems in an exchange, offering an opportunity to hear how their respective responses unfold. It also illustrates the wider interest in voice-based AI that can handle spoken interactions without requiring users to type prompts or wait for a complete response before speaking again.
Block says its published clips are live recordings in which no one steered the model during the exchanges. Subtitles were added afterwards.
The company has also highlighted a behaviour in which Dessa raises its volume when asked to speak louder. According to the team, this behaviour was not explicitly taught.
Such examples raise research questions about how models trained on speech acquire conversational conventions and respond to cues in an interaction. They also underline the importance of distinguishing between behaviour observed in a demonstration and capabilities that have been established across a broader range of situations.
The available information does not provide a comprehensive independent evaluation of Dessa’s performance or establish how consistently it handles different speakers, accents, languages and conversational settings.
Learning from speech rather than reading text
Block says the inspiration for Dessa partly comes from how babies learn to speak. Children develop language through listening to people around them and participating in interactions, rather than relying on written text as their primary source of language learning.
The company’s researchers are exploring whether an AI model can learn useful conversational capabilities through a similar emphasis on listening. The comparison is presented as a motivation for the research, rather than evidence that an AI model learns in the same way as a child.
Training directly on audio also raises questions about what information a model can acquire from spoken exchanges. These include how it learns facts and languages, how it interprets information supplied during a conversation and how those abilities develop without relying on text-based pretraining.
Block identifies these as research questions that the project could help investigate. The company has not presented the demonstrations as definitive answers to how speech-trained AI acquires knowledge.
The distinction matters because a system that reproduces the rhythm of human speech may still have limitations in factual accuracy, contextual understanding and reasoning. Researchers will need to examine these capabilities separately from the model’s ability to sound natural.
Safety is a central motivation for the project
Alongside curiosity about language learning and human-AI interaction, Block identifies safety as one of Dessa’s main motivations.
The developers argue that an AI system’s inability to recognise distress in a person’s voice could create safety problems, particularly as AI systems become more capable and operate with greater independence.
Spoken communication can convey signals that are not always obvious from the words themselves. Changes in tone, pacing and vocal intensity may indicate that someone is distressed or struggling to communicate. A system that ignores these features could miss important context during an interaction.
Dessa’s emphasis on learning directly from speech is intended to support research into how AI systems process such information. However, the demonstrations described by Block do not, by themselves, establish that the model can reliably recognise distress or respond appropriately in safety-critical situations.
The company’s stated goal is to improve understanding between humans and AI while investigating the implications of speech-based learning. Evaluating those claims would require broader testing beyond selected audio examples.
Local operation and early access for researchers
According to the development team, Dessa can run locally on a MacBook Pro equipped with an M4 Max chip or a more powerful configuration. Local operation could allow researchers to experiment with the model on compatible hardware rather than relying entirely on a remote service.
Block says a small group of AI safety researchers will receive early access. The initial release is therefore focused on research and evaluation rather than indicating that the model is broadly available to the public.
The company has not provided, in the information accompanying the announcement, a comprehensive set of benchmark results or a detailed comparison with other speech models. Further evaluation would help clarify the system’s performance, limitations and suitability for different applications.
Early access for safety researchers may also help investigate how speech-trained systems handle ambiguous instructions, sensitive conversations and potentially distressing situations.
What Dessa could mean for voice AI
Dessa represents Block’s exploration of a different approach to conversational AI: learning directly from human speech rather than depending on text-based training and conventional speech-to-text processing.
Its demonstrations draw attention to features that make conversations feel natural, including interruptions, pauses, laughter and changes in vocal intensity. At the same time, mistaken identities and unexpected shifts in topic show that realistic delivery does not guarantee accurate understanding.
The project also raises questions about how AI systems acquire factual knowledge, learn languages and recognise emotional cues when trained primarily on audio.
Block plans to make the model available initially to a limited group of AI safety researchers, with further investigation expected to focus on its capabilities and limitations. For now, Dessa is best understood as a research effort into speech-based learning and more natural human-AI interaction, rather than a demonstration that all the challenges of conversational AI have been solved.