ElevenLabs Voice AI Platform Guide (July 2024) | Will Spurlock
Imagine speaking with an artificial intelligence assistant that laughs naturally, pauses to think, recognizes changes in your voice, and responds before you have completely finished talking. Instead of sounding like a computer reading a script, it feels more like having a spontaneous conversation with another person. 🤖🎙️
That is the experience that Block, Inc. wants to explore with its new voice AI model, Dessa.
Announced on October 23, 2026, by Block co-founder Jack Dorsey, Dessa is designed to make conversations between humans and artificial intelligence feel more natural. The company says the model was trained from scratch using human conversation audio rather than text or previously trained models. It also claims that the system can listen and speak simultaneously, allowing it to respond to conversational cues that conventional voice interfaces may miss. TechCrunch+1
The announcement represents an interesting development in voice-based AI. Most people are familiar with digital assistants that speak in carefully paced sentences, wait for a question to finish, and then produce a response. Dessa aims to move beyond that pattern by capturing more of the rhythm, emotion, and spontaneity associated with human speech.
Yet the technology is not without questions. Natural-sounding AI could make digital assistants more useful and accessible, but it could also make fraudulent calls and impersonation more convincing.
Block is initially offering Dessa as a limited research preview, with early access intended for a small group of AI safety researchers. The company has not yet announced a general public release date. CoinDesk+1
This article explores how Dessa works, what makes it different, why Jack Dorsey’s announcement matters, and what the technology could mean for businesses, consumers, and the future of conversational AI.
Related tags: Jack Dorsey, Block Inc., Dessa AI, Voice AI, Artificial Intelligence, Speech Technology, Machine Learning, AI Innovation, AI Safety, Conversational AI, Technology News
What Is Dessa, and Why Is Everyone Talking About It?
Dessa is a speech-to-speech AI model developed by a research team within Block. Its purpose is to make spoken interaction with AI feel less mechanical and more like an ordinary conversation.
The model attracted attention after Dorsey introduced it publicly and described the goal as creating a more authentic conversational experience. Rather than relying on a conventional text-based interaction pipeline, Dessa is designed to process spoken input and generate spoken output directly. TechCrunch+1
This approach changes how an AI assistant can respond to the small details of conversation.
Human communication is about more than the words people say. Tone, timing, pauses, laughter, hesitation, and changes in volume can influence the meaning of a sentence.
Consider someone saying, “I’m fine.”
Depending on the tone, that statement might communicate genuine reassurance, frustration, exhaustion, or disappointment. A system that processes only the words may miss some of these signals. A speech model designed to retain audio information may have a better opportunity to interpret them.
Dessa’s developers want to explore this possibility by training a model directly on human conversation.
Why natural speech matters
Voice technology has become an important interface for smartphones, smart speakers, customer service systems, accessibility tools, and AI assistants.
However, many conversational systems still sound noticeably different from people. They may pause at unusual moments, fail to respond naturally to interruptions, or produce speech that feels overly polished.
Dessa aims to improve these aspects by learning patterns from audio itself.
The company’s demonstrations include conversations with varied pacing, emotional expression, and spontaneous reactions. Some examples sound surprisingly natural, while others reveal awkward or unexpected responses. The Verge+1
That mixture is important. A compelling demonstration can show what a model is capable of producing, but it does not establish that every conversation will be equally natural or reliable.
For now, Dessa is best understood as an experimental step toward more fluid voice interaction, rather than a finished replacement for existing assistants.
How Dessa Works: Speech In, Speech Out 🎧
Hermes Voice Mode는 채팅을 말하는 작업 루프로 바꾼다 | DEXA Blog
One of the most distinctive aspects of Dessa is its speech-to-speech architecture.
In a conventional voice assistant, a spoken conversation may pass through several stages. The system first converts speech into text, processes the text to determine an answer, and then converts that answer into speech.
This pipeline can work well, but converting speech into text may remove information about how something was said. A transcript usually records the words while losing many details of timing, vocal delivery, and emotion.
Dessa is designed to process incoming audio and generate outgoing audio without requiring a separate text transcription stage in the middle. Block says its model was trained from scratch on human conversations rather than text or pretrained models. TechCrunch+1
What does full-duplex mean?
Dessa is described as a full-duplex speech model. In this context, full-duplex means that the system can listen and speak at the same time.
This is different from a strictly turn-based interaction in which a person must finish speaking, wait for the system to process the request, and then listen to a complete response.
In natural conversations, people often provide short acknowledgments while another person is talking. They may interrupt to clarify something, laugh in response to a joke, or begin answering before the other speaker has completely finished.
A full-duplex system is designed to handle more of these overlapping conversational patterns.
For example, a user might say, “I was thinking about going to—” and then pause. A conversational assistant could respond in a way that acknowledges the unfinished thought rather than waiting through an unnecessarily rigid sequence.
That does not mean the model always knows when an interruption is appropriate. Managing conversational timing remains a challenging problem, and the quality of its behavior needs to be tested across many different situations.
Why direct audio processing could be useful
Speech contains information that written text cannot fully represent.
A person speaking quickly may be excited, anxious, or simply in a hurry. A long pause may indicate uncertainty, emotion, or a distraction. Laughter can express happiness, embarrassment, nervousness, or sarcasm.
By learning directly from speech, a model may preserve more of these signals during the interaction.
This could make voice assistants more responsive in settings where the user’s emotional state or conversational timing matters.
However, recognizing vocal patterns is not the same as understanding a person’s internal feelings with certainty. A voice model can misinterpret sarcasm or confuse tiredness with frustration.
Dessa’s potential should therefore be evaluated in terms of observable conversational performance, not assumptions that a human-like voice means the system genuinely experiences human emotions.
🧠 Trained on Human Conversation Rather Than Text
OpenAI to rebuild how AI speaks, listens, and interrupts you – StrongYes
One of the most interesting claims behind Dessa is its training approach.
Block says the model was trained from scratch using human conversation audio alone, without text data or pretrained models. The company also said its team used a single NVIDIA GB200 NVL72 computing system during training. TechCrunch+1
This differs from the development approach commonly associated with language models, which learn patterns from large collections of text before being adapted for speech-related tasks.
Learning from how people actually speak
Real conversations are rarely as orderly as written documents.
People repeat words, change direction halfway through a sentence, speak over one another, hesitate, use informal expressions, and communicate meaning through their voices as well as their words.
A model trained on human conversations may learn statistical patterns associated with these behaviors.
For example, it might learn when a short acknowledgment is appropriate, how a speaker’s pace changes during an explanation, or how laughter can occur naturally in the middle of a sentence.
This does not mean that training on speech automatically produces a perfect conversational partner. The quality of a model depends on its architecture, training data, optimization process, and the evaluation methods used to measure its behavior.
Nevertheless, Dessa offers an interesting research direction: exploring how much conversational ability can emerge when a system learns directly from audio rather than relying on text as the primary representation.
Can an AI learn knowledge through sound alone?
This is a particularly intriguing research question.
Block’s team has suggested that Dessa developed knowledge and reasoning abilities through audio-based training. The Verge+1
If further testing supports that claim, it could help researchers better understand the relationship between language, speech, and learning.
However, important details remain undisclosed. Block has not published the model’s parameter count, comprehensive benchmark results, or enough independent evaluation data to establish exactly how its capabilities compare with other systems. The Indian Express+1
That means claims about its intelligence should be treated as preliminary.
A voice model that can hold an entertaining conversation may still struggle with factual accuracy, complicated instructions, mathematical reasoning, or unfamiliar topics.
The most meaningful assessment will come from repeatable testing across a broad range of tasks, not simply from a few impressive audio demonstrations.
💻 A Voice AI Model That Can Run on a Laptop
Best Voice Recognition Software for Mac [2026 Tested]
Another notable aspect of Dessa is its reported ability to run locally on consumer hardware.
Block says the model can run on a MacBook Pro equipped with an M4 Max chip or a more powerful configuration. The company has also described inference performance as faster than real time. TechCrunch+1
If these capabilities translate into practical applications, they could make voice AI more flexible and accessible.
What local AI processing means
Many AI services process requests on remote servers. When a user speaks into an application, audio may be transmitted to a cloud service, processed there, and returned as a response.
Local inference moves some or all of that processing onto the user’s own device.
This approach can offer several advantages:
- Privacy: Audio may be processed without needing to send every conversation to a remote inference service.
- Responsiveness: A well-optimized local model can reduce dependence on network latency.
- Availability: Some applications may continue working without a constant internet connection.
- Control: Developers can have greater control over deployment and integration.
These advantages are not automatic. Privacy depends on the complete application, including whether it records conversations or transmits logs. Performance depends on hardware, model size, and workload. A local model may also have fewer capabilities than a larger cloud-based system.
Still, the possibility of running a sophisticated conversational model on a personal computer is an important direction for AI development.
Why device requirements matter
Running an AI model locally requires sufficient memory, computing capacity, and software support.
Block’s stated MacBook Pro requirement suggests that Dessa is designed to run on relatively capable hardware rather than every entry-level laptop or smartphone.
The company has not yet published enough public technical detail to establish how performance scales across different devices or how much memory the model requires. The Indian Express+1
Those details will matter to developers who want to build applications around the model.
For example, a business deploying voice assistants across hundreds of devices will need to understand the cost of hardware, installation, updates, and ongoing maintenance. A consumer application will need to balance responsiveness, battery use, and local resource consumption.
Until independent performance measurements become available, local execution should be viewed as a promising announced capability rather than a guarantee that the model will run smoothly on every supported device.
🎙️ What Makes Dessa Sound So Human?
How voice cloning is changing corporate training — Twinvoice blog
The most attention-grabbing feature of Dessa is its conversational style.
In public demonstrations, the model exhibits pauses, laughter, changes in delivery, and responses that can sound emotionally expressive. Some listeners have described the resulting voice as strikingly human-like, while others have pointed out awkward moments and unexpected changes of topic. TechCrunch+1
Several characteristics help explain why the demonstrations attracted attention.
Natural pauses and timing
A human conversation includes small gaps that communicate meaning.
A pause can signal uncertainty, give another person time to respond, or emphasize an important point. When an AI voice produces speech at a perfectly uniform pace, it may sound artificial even if the words are correct.
Dessa aims to reproduce more varied conversational timing.
Laughter and spontaneous reactions
Laughter can make conversations feel less formal. It can also communicate surprise, amusement, nervousness, or social connection.
Dessa’s demonstrations include laughter and other vocal reactions that occur within conversations. These details can make the interaction feel more spontaneous than a conventional text-to-speech system.
However, realistic laughter is not proof that a system understands a joke or experiences amusement. It is a generated vocal behavior that must be interpreted in context.
Emotional variation
Human voices change depending on the situation. A person explaining an exciting discovery may speak differently from someone discussing a stressful experience.
Dessa aims to capture more of this variation instead of delivering every sentence with the same vocal style.
That could make voice assistants feel less monotonous and potentially help them respond more appropriately when a user sounds distressed.
At the same time, emotional expression can create a misleading impression of empathy. A system can produce reassuring words and convincing vocal cues without having human feelings or personal experiences.
Users should be able to benefit from natural conversation while understanding that the technology remains an AI system.
The uncanny side of realistic speech
Human-like speech does not guarantee a coherent conversation.
Reports of Dessa’s demonstrations have described moments when it misidentifies a speaker or unexpectedly shifts topics. The Verge+1
These imperfections reveal a broader challenge: voice quality and conversational intelligence are separate dimensions.
A model might sound natural while misunderstanding a question. Another system might provide accurate information but deliver it in a noticeably robotic voice.
A genuinely useful assistant needs both reliable understanding and an appropriate conversational style.
⚠️ The Dark Side of Human-Like AI Voices
AI Voice Clone Impersonates Darktrace CEO in Deepfake Attempt
As voice AI becomes more realistic, its benefits come with a corresponding security challenge.
A voice that sounds natural can make an assistant easier to use. It can also make it harder for listeners to recognize when they are interacting with an automated system.
This concern is especially relevant to telephone scams, impersonation, and social engineering.
Industry reactions to Dessa’s announcement have reflected both excitement about its conversational potential and concern about how convincing synthetic voices could be misused. CoinDesk+1
The risk of voice impersonation
Fraudsters have already used synthetic audio and other impersonation techniques to make deceptive calls more convincing. As voice generation improves, the public may find it increasingly difficult to determine whether an unfamiliar voice is genuine.
Potential abuses include:
- Impersonating a family member or colleague during a fraudulent call.
- Pretending to represent a bank or government agency.
- Using generated voices to support deceptive customer-service interactions.
- Creating misleading audio recordings that appear to contain statements from real people.
These are general risks associated with realistic voice technology, not evidence that Block or Dessa has been used in any of these ways.
The appropriate response is to develop safeguards alongside the technology.
Why AI disclosure matters
People should be able to understand when they are speaking with an AI system, particularly in situations involving financial decisions, sensitive personal information, or emotional support.
Clear disclosure can help users make informed choices about the interaction.
Businesses deploying voice agents should also ensure that their systems do not falsely claim to be human employees. Where required by applicable law, they should comply with rules governing automated calls, recording, privacy, and consent.
Practical ways to protect yourself
Consumers can reduce the risk of voice impersonation by following a few simple practices.
- Verify unexpected requests for money through a separate, trusted communication channel.
- Avoid sharing passwords, authentication codes, or financial information during unsolicited calls.
- Use a known telephone number to contact a person or organization independently.
- Treat urgency and pressure as warning signs, even when a caller sounds familiar.
- Follow any applicable guidance from banks, employers, and public authorities about verifying identities.
A convincing voice is not proof of identity. This principle becomes more important as synthetic speech becomes increasingly natural.
🏢 How Businesses Could Use Dessa
Smiling female customer support operator pointing at a computer screen while working at her desk
If Dessa becomes broadly available and its capabilities hold up under independent testing, it could support several categories of business applications.
Its local-processing potential and speech-to-speech design are particularly interesting for organizations seeking faster, more conversational interfaces.
Customer service and virtual receptionists
Businesses could use natural-sounding AI to answer routine questions, explain product features, or guide customers through basic troubleshooting.
A system capable of handling interruptions and responding to conversational cues could make these interactions less rigid.
However, businesses would still need safeguards to ensure that agents provide accurate information, disclose that they are automated, and transfer complicated or sensitive cases to human employees.
Accessibility and assistive technology
Voice AI may help people interact with digital systems when typing is difficult, inconvenient, or inaccessible.
Potential applications include hands-free interfaces, spoken navigation, conversational reading assistance, and voice-driven productivity tools.
A more natural conversational rhythm could make these systems easier to use, although accessibility benefits would need to be evaluated with users who have different speech patterns, accents, hearing abilities, and communication needs.
Education and language learning
Conversational voice systems could provide students with opportunities to practice spoken language, ask questions, and receive immediate feedback.
A system that handles interruptions and conversational pauses may be better suited to open-ended speaking practice than a rigid question-and-answer interface.
However, educational deployments should be assessed for factual reliability, age-appropriate behavior, privacy, and the risk of presenting generated information as authoritative instruction.
Personal productivity
Voice agents may eventually help users organize tasks, summarize information, brainstorm ideas, and interact with applications without constantly switching between keyboards and screens.
A natural voice interface could make these interactions feel more fluid.
Still, any agent connected to email, calendars, payment systems, or workplace databases needs carefully controlled permissions. A conversational interface should not be allowed to perform sensitive actions merely because it can imitate a human conversation.
🔬 Why AI Safety Researchers Are Being Invited First
Dessa is currently a research preview rather than a generally available product. Block has said it is inviting a small group of AI safety researchers to examine the model. CoinDesk+1
This limited approach is significant because the risks of realistic voice AI cannot be fully understood through public demonstrations alone.
Researchers need opportunities to examine how a system behaves across varied scenarios, including situations that may reveal weaknesses.
Testing for emotional misinterpretation
A voice model should be evaluated on whether it responds appropriately when people sound confused, distressed, frustrated, or uncertain.
For example, a user reporting severe pain should not receive a casual response simply because the model fails to recognize the seriousness of the situation.
At the same time, vocal cues can be ambiguous. A model should avoid treating its interpretation of someone’s emotional state as a certain fact.
Testing should measure both its ability to recognize relevant signals and its tendency to make incorrect assumptions.
Testing reliability across different speakers
People differ in accent, speaking speed, vocal range, language background, and communication style.
A model that performs well with a narrow group of speakers may struggle with others. Safety research should therefore examine how reliably the system works across diverse users and noisy environments.
Testing impersonation risks
Researchers should also evaluate how easily the model can be used to deceive listeners, imitate identifiable people, or produce misleading audio.
Possible safeguards include consent requirements, limits on identity imitation, clear AI disclosure, and systems that help recipients verify the origin of audio.
The precise safeguards will depend on the model’s capabilities and intended applications.
Independent evidence matters
Block’s demonstrations provide an initial view of Dessa’s potential, but independent testing is necessary to determine how the model performs outside carefully selected examples.
Researchers should examine conversational quality, response latency, factual reliability, privacy behavior, and misuse resistance.
Until detailed results are published, claims that Dessa is the most natural or advanced voice model should be understood as company positioning rather than a settled scientific conclusion.
🌐 The Future of Voice AI: What Comes Next?
Glas kao budućnost AI: ElevenLabs vizija interfejsa – agenti.rs
Dessa arrives during a period of rapid progress in conversational AI. Companies are exploring systems that can communicate through speech, interpret context, and respond with fewer delays and less rigid turn-taking.
The larger shift is from treating voice as an optional feature to treating it as a primary interface for interacting with intelligent software.
For decades, people have communicated with computers mainly through keyboards, screens, menus, and buttons. Voice interfaces offer another possibility: a user can simply speak, interrupt, clarify, and continue a conversation without navigating a complicated interface.
Will voice become the main interface for AI?
Voice could become especially valuable when people need hands-free assistance or when a spoken exchange is more convenient than typing.
Drivers, customer-service employees, educators, healthcare administrators, and people using accessibility tools may all benefit from more responsive speech interfaces.
However, voice will not replace every other form of interaction. Written communication remains valuable when precision, privacy, searchability, or detailed review matters.
The most useful AI systems may combine voice, text, visual information, and conventional application controls rather than relying on a single interface.
Can local AI change the economics of voice technology?
Running models on personal devices could reduce reliance on expensive cloud infrastructure for some applications. It may also offer organizations greater control over sensitive audio.
But the economics depend on hardware costs, energy use, model efficiency, maintenance, and the level of performance users expect.
Local processing may be attractive for some applications, while cloud-based systems may remain preferable for tasks that require larger models or centralized management.
Dessa’s reported ability to run on capable consumer hardware makes it an interesting development, but its real-world advantages will become clearer only when independent performance and resource measurements are available. The Indian Express+1
The importance of trust
The success of voice AI will depend on more than whether a model sounds human.
Users need systems that understand requests, respect privacy, communicate limitations, and behave predictably. Organizations need reliable tools for monitoring, updating, and securing these systems.
A realistic voice can make an assistant more approachable, but it can also increase the risk of misplaced trust. Developers should therefore prioritize transparency and security alongside conversational quality.
The strongest voice AI products will not simply imitate the rhythm of human speech. They will combine natural communication with dependable behavior and clear boundaries.
Conclusion: A Promising Step Toward More Natural AI Conversations
Jack Dorsey’s Block has introduced an intriguing direction for voice technology with Dessa, a model designed to learn directly from human conversation and support simultaneous listening and speaking.
Its reported ability to run locally on a MacBook Pro with an M4 Max chip or better could make the technology interesting to developers exploring responsive, device-based voice applications. Meanwhile, its speech-to-speech architecture raises important questions about how AI can learn from vocal timing, emotion, and conversational context. TechCrunch+1
Nevertheless, the technology remains in a limited research preview. Public demonstrations are promising, but they do not yet establish its overall performance, reliability, or safety compared with competing models.
The opportunities are substantial. Natural voice interfaces could improve customer service, accessibility, education, and personal productivity. The risks are equally important, particularly around impersonation, privacy, and the tendency to trust voices that sound human.
The next stage of development should focus on independent evaluation, transparent disclosure, and safeguards against misuse.
