Can an AI hold a natural conversation without ever reading a word? Block, the financial-services company co-founded by Jack Dorsey, introduced Dessa on October 23, 2026, as an experimental speech-to-speech model trained exclusively on human conversation. The system is designed to listen and speak at the same time, preserve vocal nuance, and explore whether conversational intelligence can emerge directly from sound.
| Developer | Block, the company co-founded by Jack Dorsey |
|---|---|
| Announcement | October 23, 2026 |
| Model type | Full-duplex, direct speech-to-speech AI |
| Training | Human conversational audio; no text or pretrained models, according to Block |
| Local hardware | MacBook Pro with an Apple M4 Max chip or better |
| Availability | Research preview with restricted early access and a waitlist |
Key takeaways
- Dessa converts audio directly into audio instead of routing conversations through a conventional text layer.
- Block says it trained the model from scratch using only human conversation, without text or pretrained models.
- The system is not yet a commercial product, and its public accuracy, scalability, pricing, API, and model size remain undisclosed.
What makes Dessa different?
Most voice assistants follow a pipeline: they transcribe speech into text, send that text to a language model, and convert the response back into speech. Block says Dessa takes a different approach. It receives audio and produces audio directly, avoiding an intermediate transcription step that can flatten emotion, hesitation, emphasis, laughter, and timing.
Dessa is also full-duplex, meaning it can listen and speak simultaneously. In Block’s demonstrations, the model reportedly handles interruptions, pauses, laughter, backchannel responses such as “yeah” and “got it,” and changes in vocal tone. Block describes the project as a way to investigate whether an AI can learn conversational behavior directly from spoken interaction.
- Direct audio input and audio output
- Simultaneous listening and speaking
- Recognition of conversational timing and interruptions
- Responses to laughter, tone, pauses, and other vocal cues
- Inference that runs faster than real time, according to Block
- Local operation on a MacBook Pro with an M4 Max chip or better
How was Dessa trained?
Block says Dessa was trained from scratch on nothing but human conversation. The company says the model used no written text and no pretrained models, making its development distinct from the text-heavy training approach associated with many large language models.
The company compares the research direction, in part, with how infants acquire language: by listening to spoken interaction rather than reading. Block says this creates an opportunity to study how efficiently a speech-native model can learn language, facts, accents, pauses, breathing, sarcasm, humor, inflection, emotion, and conversational timing.
We trained Dessa from scratch on nothing but human conversation.
Block’s official Dessa project page
Why it matters
Good to know: Dessa is best understood as a research prototype testing speech-native AI, not as a finished voice assistant or broadly available Block product.
The project challenges a common assumption in voice AI: that spoken interaction should first be reduced to text. A transcription layer can capture the words someone says while losing information about how those words are delivered. For example, the same sentence can carry different meanings when the speaker changes emphasis, pauses, or vocal tone.
That difference could affect accessibility, education, and voice interfaces. A system that recognizes emotional and vocal signals may respond more appropriately in situations where a user’s condition is conveyed through delivery rather than wording. Block specifically identifies the possibility of an AI continuing without recognizing that someone is in serious pain as a safety concern.
- October 23, 2026
Block introduces Dessa as a full-duplex speech model trained from scratch on human conversation.
- October 24, 2026
Reports highlight Dessa’s local operation, human-like vocal behavior, and concerns about possible misuse.
- October 25, 2026
The latest reported information continues to describe Dessa as a restricted research preview with no publicly disclosed model size, API, pricing, or standardized benchmark results.
What to watch
- Whether Dessa becomes available beyond the initial group of AI-safety researchers and waitlist participants.
- Independent evaluations of factual reliability, conversational consistency, latency, and performance across different hardware.
- Whether the speech-only approach can outperform established voice-AI systems outside curated demonstrations.
- How Block addresses impersonation, social engineering, phone-based fraud, and other misuse risks associated with realistic voices.
- Whether local processing improves privacy by reducing cloud transmission while making centralized abuse detection more difficult.
- Future disclosures about model size, training-data governance, API access, pricing, and standardized evaluations.
The demonstrations suggest that natural-sounding interaction does not automatically equal broad reliability. Reports describe moments in which Dessa responds awkwardly, misunderstands a situation, or makes conversational mistakes. The distinction matters: a system may sound emotionally convincing while still requiring significant evaluation for facts, reasoning, and dependable task performance.
Availability is also limited. Block is prioritizing a small group of AI-safety researchers for early access, while other interested users can join a waitlist. The company has not publicly announced model weights, an API, pricing, a parameter count, or standardized benchmark results.
Heads up: The same realism that could make voice interfaces more useful could also make impersonation, social engineering, and phone-based fraud more convincing.
Dessa’s development is led by Ragavan Thurairatnam within a small AI research group at Block. The team includes former members of Dessa, the Toronto deep-learning startup, but the new model is a Block project. The group says its motivations include curiosity, more natural human-AI interaction, and a closer examination of safety issues involving vocal and emotional signals.
For now, Dessa’s significance lies less in immediate product availability than in the direction it represents. By training a model on human speech alone and allowing it to listen and speak concurrently, Block is testing a different path for conversational AI. The next phase will show whether that path can deliver reliable, scalable, and safe performance beyond demonstrations.