Dessa is an experimental speech-to-speech AI model developed by Block, the financial technology company led by Jack Dorsey, who announced the project on October 23, 2026. The model is designed to listen and speak simultaneously, respond to interruptions and preserve conversational cues such as laughter, pauses, sarcasm and changes in tone. Block says it was trained from scratch using human conversation audio rather than text or pretrained models.
The project represents a different approach to voice AI. Many existing systems convert speech into text, process the text using a language model and then generate spoken output. Dessa instead aims to learn conversational behavior directly from audio. Block says the model can run locally on a MacBook Pro with an M4 Max chip or better, but it remains a research preview, with early access planned for a small group of AI safety researchers.
Key Takeaways
- New AI model: Jack Dorsey’s Block announced Dessa on October 23, 2026.
- Full-duplex conversations: The model can listen and speak simultaneously, enabling interruptions and more natural conversational timing.
- Speech-only training: Block says Dessa was trained from scratch on human conversation audio without text or pretrained models.
- Local execution: The company says it can run on a MacBook Pro with an M4 Max chip or better.
- Human-like expression: Demonstrations feature laughter, emotional changes, pauses and different speaking styles.
- Limited access: Dessa is currently a research preview, not a publicly available product with an announced general release date.
- Safety questions: More natural synthetic voices could improve accessibility and interaction but may also increase the risk of impersonation and voice-based scams.
What Is Dessa, and Who Built It?
Dessa is a speech-to-speech model developed by a small research team inside Block, the company behind Square and Cash App. The team works on conversational AI and reports directly to Jack Dorsey, according to launch coverage.
The project focuses on making voice interaction more natural. Rather than treating speech as a string of words, the system aims to learn from the broader information present in human conversation, including timing, delivery and emotional expression.
Block argues that text-based processing can discard useful information. When someone speaks, their meaning may depend on whether they sound confused, excited, sarcastic or distressed—not just on the words they use.
Dessa is designed to retain more of those signals by processing audio directly. In demonstrations shared by the company and its researchers, the model responds with different vocal styles, pauses, laughter and conversational acknowledgements.
However, a natural-sounding voice does not automatically mean a system understands emotions as humans do. Its ability to interpret subtle cues and respond reliably will require broader independent evaluation.
How Full-Duplex Speech AI Works
The term full-duplex describes a system that can listen and speak at the same time. In ordinary human conversation, people regularly interrupt one another, acknowledge what they hear and begin responding before the other person has completely finished.
Many voice assistants follow a more sequential process: listen to the user, process the request, generate an answer and then speak. That approach can introduce pauses and make interruptions awkward.
Dessa is designed to support overlapping interaction instead.
| Feature | Conventional sequential voice pipeline | Dessa’s stated approach |
|---|---|---|
| Input processing | Often converts speech into text | Processes audio directly |
| Response generation | Generates text before speech in many systems | Produces speech from audio input |
| Interruptions | May require turn-taking or interruption handling | Designed to listen while speaking |
| Emotional cues | May rely on audio analysis or text-based labels | Aims to learn cues directly from conversation audio |
| Conversational acknowledgements | Often explicitly engineered into the interaction | Designed to learn natural timing and acknowledgements |
This does not mean every competing system is strictly sequential. Other modern voice models also support low-latency, real-time conversations and interruptions. Dessa’s distinctive claim is its training approach: learning directly from speech without text or pretrained models.
Trained From Scratch on Human Speech
One of the project’s most notable claims is that Dessa learned from human conversation audio alone.
According to Block, the model was trained from scratch without text-based training data or pretrained models. The company says it learned conversational patterns from audio using a single NVIDIA GB200 NVL72 system.
The approach raises an important research question: how much linguistic and conversational capability can a model acquire from spoken interactions alone?
Text-based language models learn from enormous collections of written material. Speech-only training places greater emphasis on vocal patterns, timing and the relationships between sounds and conversational responses.
Block says Dessa learned language from hundreds of times fewer words than large language models typically require. That is a company claim, and the public information does not yet include sufficient independent benchmarks to establish how its knowledge, reasoning or language abilities compare with other systems.
The model’s development could nevertheless help researchers investigate alternative ways of building conversational AI. A system trained directly on audio may preserve information that is difficult to represent fully through text alone.
What the Demonstrations Reveal
Block has released recordings intended to demonstrate Dessa’s conversational behavior. These include conversations featuring different voices, changing emotional tones and interactions in which the model responds to surprising information.
One example features a discussion about a surprise trip to Paris, with Dessa responding in a way that sounds emotionally affected. Other clips demonstrate laughter, natural pauses and changes in speaking style.
The demonstrations also reveal limitations. Reporting by The Verge describes moments when the model misidentifies the person speaking or moves unexpectedly to another topic. These examples suggest that realistic speech generation does not eliminate errors in conversation.
The distinction between the demonstrations and a comprehensive evaluation is important. The recordings show selected interactions, not a systematic measurement of reliability across accents, languages, noisy environments or sensitive situations.
Independent testing will be needed to determine how consistently Dessa handles interruptions, understands speakers and responds appropriately when conversations become difficult or ambiguous.
Can Dessa Run on a MacBook?
Block says Dessa can run locally on a MacBook Pro equipped with an M4 Max chip or better. Local execution could allow voice interactions to take place without sending every conversation to a remote inference service.
This could have several advantages:
- Privacy: Local processing may reduce the need to transmit audio to an external server, depending on the final implementation.
- Latency: Running on-device could reduce network delays and support responsive conversations.
- Availability: Local operation could allow some features to work without a continuous cloud connection.
- Cost control: Developers may avoid certain per-request cloud inference charges, although hardware and energy costs still apply.
These are potential advantages rather than guarantees. Actual privacy depends on what the application records or transmits, while performance depends on hardware, model configuration and workload.
Block has not yet published all the technical details developers would need to evaluate model size, memory requirements, licensing and deployment options.
Availability, Pricing and Developer Access
As of October 26, 2026, Dessa remains a research preview. Block has invited a small group of AI safety researchers to obtain early access and investigate risks the company may not have anticipated.
The company has not announced a general release date, public API, subscription pricing or downloadable model weights. Developers therefore cannot yet assume they can integrate Dessa into commercial applications.
For now, the available demonstrations provide an initial view of the technology. A broader release, technical documentation and independent benchmarks would help establish how it performs in real-world environments.
Could Dessa Make Voice AI More Useful?
More natural speech interaction could benefit applications where timing and emotional cues matter.
In customer service, a voice agent could respond more fluidly to interruptions. Accessibility tools could offer more conversational interactions for people who prefer speaking over typing. Educational applications could experiment with spoken practice that feels less like a sequence of recorded prompts.
Block’s existing business interests also provide potential application areas. Square already offers voice-based ordering tools, and improved speech models could eventually support a broader range of conversational services. However, Block has not announced a specific commercial rollout of Dessa.
The technology could also help developers build assistants that operate locally, particularly when privacy, latency or network availability is important.
The size of the opportunity will depend on whether the model can maintain accurate responses and reliable behavior beyond carefully selected demonstrations.
Safety Concerns: Realistic Voices and Potential Misuse
More natural synthetic speech presents a security challenge as well as a product opportunity.
Voice-based scams already exploit people’s trust in familiar voices. If AI systems become better at reproducing natural timing, emotion and conversational reactions, it may become harder for listeners to distinguish some generated audio from genuine speech.
Dessa’s announcement explicitly includes a safety motivation. Block says it wants outside researchers to investigate the risks of a model that learns from sound alone before wider access becomes available.
Potential areas for testing include whether the model can be used to impersonate people, how it handles sensitive personal information, whether it responds appropriately to signs of distress and what safeguards are needed for deployment.
These are questions to investigate, not proof that Dessa has been used in scams. The limited preview means the model’s real-world misuse risks and defenses remain uncertain.
The Bigger Picture
Dessa represents an alternative direction for conversational AI: instead of treating speech primarily as text that must be processed and spoken back, it attempts to learn directly from human audio. If this approach proves effective, it could influence how future voice assistants handle timing, interruptions and emotional cues.
The launch also underscores a broader trade-off in AI development. More realistic interactions can make technology easier to use, but they can also make automated systems harder to distinguish from people. Independent testing, transparent technical documentation and safeguards against impersonation will be important as the model moves beyond a research preview.
FAQs
1. What is Dessa AI?
Dessa is an experimental speech-to-speech AI model developed by Block, the company led by Jack Dorsey. It is designed to listen and speak simultaneously and learn conversational behavior directly from human speech.
2. How is Dessa different from other voice AI models?
Block says Dessa was trained from scratch on human conversation audio without text or pretrained models. It is also designed for full-duplex interaction, allowing it to listen while speaking. Other modern voice systems may support similar real-time conversational features but use different architectures and training methods.
3. Can Dessa run locally on a laptop?
Block says the model can run on a MacBook Pro with an M4 Max chip or better. Detailed information about model size, hardware requirements and performance has not yet been published.
4. Is Dessa available to the public?
No general public release has been announced as of October 26, 2026. Block is offering early access to a small group of AI safety researchers.
5. Could Dessa be used for voice scams?
Realistic synthetic speech can potentially be misused for impersonation, but the announcement does not establish that Dessa has been used in scams. Safety testing will be important to understand the risks and develop appropriate safeguards.
Looking Ahead
The next major developments to watch are independent evaluations, technical documentation and any expansion of access beyond the initial research preview. These will help determine whether speech-only training delivers practical advantages in conversational quality, latency and local deployment.
For Block, Dessa is an experiment in how people may interact with AI in the future. Its long-term significance will depend not only on how human its voice sounds, but also on whether it can communicate accurately, protect users and behave reliably outside controlled demonstrations.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
