Just imagine you’re interacting with an AI that casually responds, laughs naturally, reacts to interruptions, and modulates its tone during the conversation. Would you be able to differentiate it from another person? This is the question that has caught the attention of many following the game-changing announcement of Dessa, the latest voice AI model from Block. As AI continues to evolve, businesses offering ChatGPT Ads Services and AI SEO services are also keeping an eye on emerging technologies that could shape the future of digital interactions.
This question might’ve made you curious as well, and if you’re indeed wondering, “What is Dessa AI?”, it is an experimental speech model in development that's designed to make conversations with AI sound more natural and human.
Dessa was announced on October 23, 2026, and has generated significant interest because it learns from human conversation rather than relying on text-based training. Want to know what sets Dessa apart? This blog explores everything you need to know about Block’s latest voice AI model.
What Is Dessa AI?
Dessa is a cutting-edge, full-duplex, speech-to-speech AI model developed by an in-house team at Block, a massive fintech powerhouse led by Twitter’s co-founder Jack Dorsey. As the company behind Square and Cash App, Block has engineered Dessa from scratch, utilizing human conversation audio, without text or pretrained models.
The AI is currently in the research preview stage and is not yet available for the general public. The capabilities of the AI model have been shown through multiple demo recordings and demonstrations, but comprehensive evaluations and independent performance benchmarks by professional researchers haven’t been published yet.
How Does Dessa AI Work?
Full-Duplex Conversations
Conventional voice assistants have a uniform process - the user speaks or presents a query, the system processes the input, and the AI responds.
But Dessa is designed to eliminate the intermediate text-processing step by speaking to the user in a more natural and conversational manner. It is designed to listen and speak simultaneously. This allows the AI model to respond to interruptions and produce brief acknowledgements, as well as express emotion whilst someone is talking to it.
By making conversations more natural and fluid, this approach could change how interactions with AI models work, as neither participant has to wait for the other to finish speaking before responding.
Training on Human Speech
Dessa is exclusively trained on human conversation audio without any text or pretrained models. Block says that being trained on audio allows the model to learn how to retain conversational nuances such as intonation, pauses, and feelings.
Many traditional voice systems convert the input audio to text, after which they generate a response and convert it back to audio. As a result, some of the vocal information might get lost in between if not preserved properly.
Dessa’s functional aim is to interact directly through audio without an intermediate step to convert the query into text. However, its system architecture and performance need to be independently evaluated.
Understanding the difference
- Conventional pipeline: Starts with a user speech, the text gets transcribed, the AI comprehends the data, and generates a speech in response to it.
- Dessa's approach: The user inputs their speech, Dessa instantly processes the speech, and responds back with a speech; both listening and speaking happen at the same time.
Dessa AI vs. the Original Dessa and RealTalk
The name Dessa refers to two related but distinct developments in AI.
The original Dessa was a machine learning startup founded in 2016 and based in Toronto. They created RealTalk in 2019 as an example of how text can be converted into speech that sounds like famous podcaster Joe Rogan. In February 2020, the startup was acquired by Square, now famously known as Block. The team behind this project are now involved with the AI research at Block. The key difference is the technology's purpose:
- Original Dessa: The company created machine learning technology, including research into text-to-speech.
- RealTalk: A synthetic speech model that reads in text to create speech. It demonstrated how realistic speech cloning can become.
- Block's Dessa: Focuses on live, interactive speech conversations, with simultaneous listening and speaking.
The shared name reflects a connection through the research team, but the new model is not just an updated version of the 2019 example.
What Can Dessa AI Do?
Block's demonstrations show Dessa's expressive speech capabilities:
- Laugh naturally: Rather than speaking in a flat tone throughout the interactive session, the model has shown the ability to laugh during a sentence when apt.
- Adjust vocal intensity: In one of the demos, the model raised its volume when requested to raise its volume.
- React with surprise: It can produce audible reactions, including gasps.
- Express emotion: The human-like element is noticeable here, as Dessa has displayed its ability to manipulate its vocal delivery according to the context of a conversation. For instance, it can sound tired when discussing tiredness.
These matter because natural-sounding speech does not necessarily mean accurate reasoning or reliable answers.
These examples demonstrate that speech generation is capable of communicating more than words alone. But reports have noted that there are issues with the speech generated, such as it having filler words, false starts, corrections, and sometimes odd answers. This matters because a natural-sounding speech doesn’t necessarily ensure accurate reasoning or correct answers to user queries.
Availability and Hardware
As of now, Dessa remains in the research preview stage and has not yet been released to the public as a ready application. Block says the model can run locally on a MacBook Pro with an M4 Max chip or better. Local execution could offer practical advantages, including reduced dependence on remote servers, but actual performance and resource requirements need further testing.
Access to the AI model is currently limited to a small circle of AI safety researchers, with no confirmed public release date. The company has also kept the model's parameter count, published benchmarks, pricing, and broader distribution details under wraps.
More information about the model can be found on Block’s official Dessa page. This allows you to view demos and keep up to date with any updates. You can check the page for an active waitlist before attempting to register; the availability of a public signup has not been confirmed.
Safety Concerns
The perks of a human-like voice AI model are notable, but it also comes with certain risks that could be dangerous in the wrong hands.
Block argues that AI voice systems that are unable to properly understand vocal emotions may miss out on important signals such as discomfort or distress. It’s very important that AI models develop proper speech understanding to ensure they are capable of responding more appropriately in sensitive conversations. Although these capabilities provide great advantages, they require further testing and evaluations.
However, the problem with a human-voiced AI is that it can make it harder to differentiate an AI from a genuine person and identify impersonation or scam calls. Even familiar voices alone won’t be sufficient to prove one’s identity.
How to protect yourself from fake voice calls:
- Create a safe word that is known to only a few people. This can help verify unusual or suspicious requests.
- Call the person back using a number you already trust.
- Be cautious of urgent requests for money, passwords, or verification codes.
- Cross-check the authenticity of important requests through a different means of communication.
These precautions can be highly useful, irrespective of the voice-generation technology that is being used.
Conclusion
Block’s Dessa AI is an experimental real-time listening and responding voice model designed to make interactions feel more genuine and humanlike. The demonstrations that have been revealed by the company show promise in terms of conversational expression, but the model still requires independent testing for a comprehensive understanding of how accurate, reliable, and safe it is.
For now, Dessa remains a research preview, meaning it’s not yet available to the general public, and no confirmed release date has been announced. But considering the potential the model holds, it could become a convenient tool in the future, making it an impressive piece of tech to keep an eye on. The demos have only shown a fraction of its capabilities, leaving plenty of room to explore what this technology could achieve in the future. As AI continues to evolve, investing in AI SEO services can help businesses stay visible and adapt to emerging trends in AI-driven search.
