Block announced the launch of a new voice model named Dessa. The company released this model from its research lab. It runs locally on a consumer computer. The model sounds strangely human to listeners. It includes expressions like laughs, pauses, and cries. Jack Dorsey, the founder of Block, shared this news on social media. He stated that the model is fully trained on human speech. He also noted that it uses a single NVIDIA GPU for training. This release marks a shift in how companies build voice assistants.
Block invited safety researchers to test Dessa before anyone else could use it. The company wants to find potential risks in the model. Safety researchers will check if the bot tricks people into giving up personal data. They will also test if the bot mimics emotions too convincingly. This step is crucial for building trust in AI voice assistants. Many users worry about privacy when talking to machines. Testing with experts ensures the model does not cause harm.
Dessa differs significantly from other popular voice models like ElevenLabs. Most competitors use voice tags to drive emotion in speech generation. These tags tell the model which emotion to display at specific moments. Block claims that emotion is built directly into Dessa. It does not rely on external text instructions to steer feelings. Current models often sound unnatural because of these text-driven pipelines. Users frequently find them annoying or robotic in their delivery.
Dessa is a full-duplex speech-to-speech model. This means it can talk and listen at the same time. It does not convert speech to text and back again. This process preserves the emotional nuance of the original voice. The model runs on a single NVIDIA GPU for training. It can also run on consumer-grade hardware like a MacBook M4 Max. This hardware requirement makes it accessible for local deployment. Users do not need massive cloud servers to run it.
The technical details show how Dessa was trained entirely on speech data. Traditional models often mix text and speech during the training phase. Dessa avoids this hybrid approach to maintain authenticity. The company believes voice is the most natural way to interact with intelligence. They created Dessa to be the most authentic full-duplex speech model ever built. The model captures natural cadence and emotional pauses.
Why local voice models matter for cost and privacy is a key point. Running a model locally keeps data on the user's device. This reduces the need to send voice recordings to the cloud. It lowers the cost of sending data over the internet. Users also gain better privacy control over their conversations. Block can power more business experiences at a lower cost. Square already uses voice AI to handle ordering for businesses. Local models allow smaller businesses to adopt voice technology easily.
How OpenSmartRoute helps manage voice agent deployments is relevant here. A team using OpenSmartRoute can add Dessa to its catalogue. The router scores every candidate on quality, cost, speed, and safety. The team sets the weights for each request type. Hard rules can pin requests to on-premises models if needed. This ensures sensitive voice data stays local when required. The input guard spots prompt injection before it leaves the device.
What to do when a new voice model launches involves checking the safety record. Readers should look for safety researcher feedback before adopting the model. They can compare Dessa against existing tools like ElevenLabs v4. Testing the model with their own prompts reveals real-world performance. A savings ledger shows what each routed request cost next to the most expensive option. This helps teams make informed decisions about model selection.
The blog post mentions that current voice AI can just plain suck. It turns voice into text and throws away incredible information. Text alone cannot carry the full emotional context of a conversation. The most powerful AIs in the world cannot hear how you feel. Block argues this leads to a worse experience and potential danger. They believe this limitation is ridiculous and needs fixing.
Researchers posted multiple examples of conversations with Dessa. The agent had different voices, cadences, and emotions in these tests. For instance, the bot sounded tired when it said it was tired from work. This consistency in emotional expression is a major breakthrough. It makes the interaction feel much more like talking with a person.
Another example shows a Block team member playing a guessing game with Dessa. The model guessed the first item right straight away. The second guess did not take long either. This speed and accuracy in conversation are impressive for a local model. It demonstrates the model's ability to engage in dynamic interactions.
Dessa represents a GPT3 moment for speech technology. This comparison highlights the significant leap forward in AI capabilities. Just as GPT3 changed text generation, Dessa changes voice interaction. It allows for realistic conversations without heavy cloud infrastructure. This shift could change how businesses deploy voice assistants.
The company stated that emotion is built into the model. It is not clear if emotion is being steered by some kind of tags. This distinction matters for how the model behaves in different contexts. Some users might prefer tag-driven control over built-in emotion. Others might prefer the more natural flow of built-in emotion.
When you purchase through links in our articles, we may earn a small commission. This doesn't affect our editorial independence. Ivan covers global consumer tech developments at TechCrunch. He is based out of India and has previously worked at publications including Huffington Post and The Next Web.
You can contact or verify outreach from Ivan by emailing [email protected] or via encrypted message at ivan.42 on Signal. Cover more ground at TechCrunch's ultimate founder bootcamp of the year by bringing your co-founder or partner for free.
Warhound raises $200M to engineer war dogs and arm them with Anduril gear. Uber puts all 30,000 employees on ChatGPT Work. Airbnb limits chores for guests, sparking backlash from hosts. Year-old AI lab Flapping Airplanes raises $1B at a $6B valuation. Tesla starts Optimus robot production in California as profit falls 43%. The internet is having a very bad cybersecurity day. People are using AI to make their smart TVs dumb again.
The release of Dessa signals a new era for local AI. Companies can now build voice assistants that run on user devices. This reduces dependency on large cloud providers. It also lowers the barrier to entry for voice AI adoption. Smaller startups can compete with big tech firms. The focus on safety and privacy is a positive trend.
Block is in the business of finance and selling its solutions to small businesses. Square already uses voice AI to handle ordering. But with its own voice model that can run on-device, it can power a lot more business and consumer experiences at a lower cost. This strategy aligns with the trend of on-premises AI deployment.
The input guard spots prompt injection and personal data before a request leaves. This feature is critical for protecting user privacy. It prevents malicious inputs from leaking sensitive information. Teams can rely on this guard to secure their voice interactions.
The hosted platform keeps a models catalogue with prices and public rankings built from real traffic. This transparency helps teams choose the best models for their needs. They can see how different models perform in real-world scenarios. The catalogue is updated regularly with new models and data.
A new model is one catalogue entry and competes on the next request. Nothing else changes in the app when a new model is added. This ease of integration is a major advantage for teams. They can test new models quickly without complex setup.
It works with any OpenAI-compatible provider, with open-weight models served locally. This flexibility allows teams to mix and match different models. They can route voice requests to the best local model available. The router ensures optimal performance for every request type.
The 'osr eval' measures routing accuracy on the team's own prompts and can fail a build when it drops. This automated testing ensures the routing logic remains reliable. Teams can catch errors before they affect users. This feature builds confidence in the routing system.
A savings ledger shows what each routed request cost next to what the most expensive model would have cost. This financial tracking helps teams optimize their spending. They can see exactly how much they save by using the right model. This data supports budget planning and cost reduction efforts.
When a new voice model launches, teams should evaluate its safety first. They should check if it has passed safety researcher tests. Reading the blog post provides initial insights into the model's capabilities. Comparing it with competitors helps form a clear opinion.
Readers can try running Dessa on their own local hardware. They can test its emotional expression and conversational flow. Checking the technical details helps them understand the hardware requirements. This hands-on approach provides the best assessment of the model.
The company said the model can run locally on a consumer-grade computer such as a MacBook M4 Max or better. This specification is important for planning hardware purchases. Teams can budget accordingly for local deployment. It avoids the need for expensive enterprise servers.
Block invited safety researchers to test it before public release. This step shows a commitment to responsible AI development. It reduces the risk of deploying a harmful model. Teams should look for similar safety vetting processes when choosing models.
Dessa is a full-duplex model, which means it can talk and listen at the same time. This capability replaces the need for separate speech-to-text and text-to-speech steps. It simplifies the architecture of voice applications. Developers can build faster and more efficient voice agents.
Typically, speech models are paired with a large language model for knowledge retrieval. Block said that Dessa's knowledge stems from the conversational data it has. This design choice reduces the need for external knowledge bases. It makes the model more self-contained and easier to deploy.
The company claims that emotion is built into its model. It is not clear if that emotion is being steered by some kind of tags. This ambiguity requires careful testing by users. They need to verify how emotion behaves in complex scenarios.
We believe voice is the most natural way to interact with intelligence, so we created Dessa. This philosophy drives the design and training of the model. It prioritizes human-like interaction over strict logical constraints. Users will appreciate the more natural feel of the voice.
Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. This description highlights the primary goal of the project. The model aims to bridge the gap between human and machine communication. It seeks to make AI interactions feel genuinely human.
Jack Dorsey's Block released a new voice model out of the company's research lab. This announcement comes from the leader of the company. It signals a strong push into consumer-facing AI products. The focus is on authentic human-like experiences.
The company said the model, called Dessa, is trained fully on human speech and a single NVIDIA GPU. These training details are critical for understanding the model's capabilities. They explain the source of the model's knowledge and its size.
Dessa is a full-duplex model, which means it can talk and listen at the same time. This technical definition is essential for understanding its functionality. It distinguishes the model from sequential speech systems.
Block said that the current models can sound unnatural and annoying because they have text in the pipeline that drives speech generation. This critique of existing models sets the stage for Dessa's introduction. It explains the problem that Dessa aims to solve.
A lot of models like ElevenLabs v4 or Fish Audio Drama 3 use voice tags to drive emotion in speech generation. These examples illustrate the current state of the market. They show the limitations of tag-driven emotion control.
Dorsey's company claims that emotion is built into its model. This claim differentiates Dessa from competitors. It suggests a more integrated approach to emotional expression.
"Interacting with current voice AI can just plain suck. It turns your voice into text and throws away an incredible amount of information text alone can't carry." This quote from the company highlights the core value proposition. It emphasizes the loss of information in text-based systems.
The most powerful AIs in the world can't hear how you feel. We think that's ridiculous, leads to a worse experience, and is potentially dangerous. This statement underscores the potential risks of current technology. It argues for a more empathetic AI approach.
The company's researchers posted multiple examples of conversations with Dessa where the agent had different voices, cadences, and emotions. These examples serve as proof of concept for the model's capabilities. They demonstrate the model's ability to mimic human behavior.
For instance, in the conversation, the bot said it was tired from work, and at that point it sounded tired. This specific example illustrates the model's emotional consistency. It shows how well the model captures human states.
We made the first AI model you can have a realistic conversation with that was trained entirely on speech! This statement emphasizes the novelty of the model. It positions Dessa as a pioneering product in the field.
Dessa is a full-duplex speech-to-speech model that you can run on consumer GPUs. This sentence summarizes the key technical and deployment features. It combines capability with accessibility.
This is a GPT3 moment for speech. This comparison places Dessa in the context of AI history. It suggests a paradigm shift similar to the GPT3 era.
In the example below, the bot was talking about the person making it comfortable while sounding emotional and almost breaking down. This example shows the model's ability to handle complex emotional scenarios. It demonstrates empathy in the voice.
Another one shows a Block team member playing a game with the voice assistant. This example shows the model's utility in interactive entertainment. It highlights the model's responsiveness and engagement.
Watch Cara play a guessing game with Dessa. The model guesses the first one right straight away, and the second doesn't take long either. This specific performance metric shows the model's speed and accuracy. It proves the model is not just slow and robotic.
The company said that the model can run locally on a consumer-grade computer such as a MacBook M4 Max or better. This specification defines the hardware requirements for deployment. It makes the model accessible to a wide range of users.
However, there is no public access to Dessa at the moment. This indicates that the model is in a closed beta phase. Access is limited to safety researchers and select partners.
Block has put out a call for safety researchers to test the model. This action shows the company's commitment to safety and quality. It invites external scrutiny of the model's performance.
This is likely to test out cases where a voice agent could trick people on the other end. This explains the purpose of the safety testing. It focuses on adversarial scenarios and potential manipulation.
Block is in the business of finance and selling its solutions to small businesses. This context explains the company's commercial focus. It highlights the practical applications of the voice model.
Square already uses voice AI to handle ordering. This example shows existing use cases for voice AI in Block's ecosystem. It demonstrates the potential for broader adoption.
But with its own voice model that can run on-device, it can power a lot more business and consumer experiences at a lower cost. This statement outlines the strategic advantage of the new model. It emphasizes cost efficiency and scalability.
When you purchase through links in our articles, we may earn a small commission. This is a standard disclosure for the publication. It clarifies the financial relationship between the site and advertisers.
Ivan covers global consumer tech developments at TechCrunch. He is based out of India and has previously worked at publications including Huffington Post and The Next Web. This information identifies the author and his background.
You can contact or verify outreach from Ivan by emailing [email protected] or via encrypted message at ivan.42 on Signal. This provides contact information for the author. It ensures secure communication channels.
Cover more ground at TechCrunch's ultimate founder bootcamp of the year by bringing your co-founder or partner for free. This is a promotional offer for the publication. It encourages community engagement and networking.
Warhound raises $200M to engineer war dogs and arm them with Anduril gear. This headline reflects the diverse range of tech news covered. It shows the breadth of the publication's interest.
Uber puts all 30,000 employees on ChatGPT Work. This headline highlights the trend of enterprise AI adoption. It shows the scale of AI integration in large companies.
Airbnb limits chores for guests, sparking backlash from hosts. This headline touches on the social impact of AI. It shows the complexities of AI implementation.
Year-old AI lab Flapping Airplanes raises $1B at a $6B valuation. This headline tracks the funding landscape for AI startups. It shows the high valuations in the sector.
Tesla starts Optimus robot production in California as profit falls 43%. This headline covers the intersection of robotics and finance. It shows the challenges of scaling new technologies.
The internet is having a very bad cybersecurity day. This headline addresses a critical security issue. It warns readers about potential threats.
People are using AI to make their smart TVs dumb again. This headline explores the unintended consequences of AI. It shows the dual-use nature of AI technology.
The release of Dessa signals a new era for local AI. Companies can now build voice assistants that run on user devices. This reduces dependency on large cloud providers. It also lowers the barrier to entry for voice AI adoption.
Smaller startups can compete with big tech firms. The focus on safety and privacy is a positive trend. Teams should look for similar safety vetting processes when choosing models.
Dessa is a full-duplex model, which means it can talk and listen at the same time. This technical definition is essential for understanding its functionality. It distinguishes the model from sequential speech systems.
Block said that the current models can sound unnatural and annoying because they have text in the pipeline that drives speech generation. This critique of existing models sets the stage for Dessa's introduction. It explains the problem that Dessa aims to solve.
A lot of models like ElevenLabs v4 or Fish Audio Drama 3 use voice tags to drive emotion in speech generation. These examples illustrate the current state of the market. They show the limitations of tag-driven emotion control.
Dorsey's company claims that emotion is built into its model. This claim differentiates Dessa from competitors. It suggests a more integrated approach to emotional expression.
How OpenSmartRoute helps
A team routing requests through OpenSmartRoute gains immediate access to Dessa without changing its catalogue. The router treats Dessa as a new catalogue entry and scores it on quality, cost, speed, and safety. It compares Dessa against existing models to find the best fit for each conversation.
The system ensures personal data stays on an on-premises model if needed. An input guard spots prompt injection before a request leaves the network. This prevents sensitive information from reaching external voice models.
The savings ledger shows what each routed request cost next to the most expensive option. Teams can set hard rules so a region or cost cap is never crossed. OpenSmartRoute learns from outcomes, giving more traffic to models that answer well.
Source: https://techcrunch.com/2026/10/24/blocks-new-voice-model-sounds-eerily-human/