Block unveils Dessa, a speech-only AI model for emotional voice cues
Dessa processes audio directly instead of converting speech to text, with Block seeking security researchers before any public release.
Block has introduced Dessa, an AI speech model that the company says was trained entirely on human speech rather than text. The Decoder reports that the model is designed to process spoken audio directly, a choice that could matter for teams building voice agents where tone, pauses and emotion change the meaning of a request.
For creators, marketers and small businesses, the news is less about an immediately available tool and more about the direction of voice AI. Block is positioning Dessa as an attempt to move beyond voice interfaces that hear words but miss how those words are said.
Dessa skips the speech-to-text step
According to Block, Dessa was built from scratch using only human speech data, with no text data in the training mix. Instead of turning a customer’s voice into text and then reasoning over that transcript, the model works with speech as audio.
Block says that direct audio processing helps Dessa pick up cues such as emotion, sarcasm, laughter and silence while a conversation is happening. Those signals often disappear or become ambiguous when speech is flattened into text. A phrase that looks neutral in a transcript can sound annoyed, joking or hesitant in the original audio.
The company also says Dessa supports full-duplex communication, meaning it can listen and speak at the same time. That is important for voice agents because real conversations involve interruptions, backchannels and overlapping speech rather than neat turn-taking.
Dessa is not publicly available. Block says its team is first looking for AI security researchers to test the model for unexpected risks, with interested researchers able to sign up through a form.
Why customer-facing voice apps should pay attention
The most obvious audience for this kind of model is anyone building voice systems that talk directly to customers: ordering, support, intake, booking or guided help. Block owns Square and Cash App, and the company’s interest in more natural voice AI sits alongside broader moves toward spoken interfaces, including Square’s AI ordering kiosk, which customers can talk to (related Emonarc coverage).
For a restaurant, clinic, shop or service business, the hard part is not only recognizing words. A useful agent may need to notice when a customer is confused, impatient, joking or correcting themselves. Block’s argument is that text-first systems can miss those cues because they depend on transcripts rather than the full acoustic signal.
The on-device detail is also notable. Block says Dessa runs locally on a MacBook Pro M4 Max. The company also says the model was trained on a single NVIDIA GB200 NVL72. If similar systems become productized, local processing could be relevant for teams that care about responsiveness or want more control over where audio is handled, though Block has not announced a product release.
How this fits into the voice AI market
Dessa arrives in a crowded field of voice AI experiments. The Decoder notes that Hume AI launched EVI, short for Empathic Voice Interface, in March 2024. Kyutai followed with Moshi in September 2024. LAION’s Open Empathic project has been working since late 2023 on empathetic speech synthesis models built with an open-source audio dataset.
OpenAI has also continued improving ChatGPT voice interactions. According to The Decoder, OpenAI overhauled ChatGPT’s Advanced Voice Mode in June 2025 and has kept refining intonation, pauses and more realistic expressions of empathy and sarcasm.
That said, the source makes an important distinction: sounding emotional is not the same as being emotional or intelligent. A realistic voice can create a strong impression, but the quality of a voice agent still depends on whether it can reason well while responding. For business users comparing voice tools, that means demos should test more than vocal warmth.
Existing voice generation tools such as ElevenLabs, Murf AI and Speechify are often used in creator and business workflows, but Dessa is being presented as a research model focused on speech understanding and interaction rather than as a public production tool. Readers comparing the broader category can also see our guide to the best AI voice generators.
What to watch before Dessa matters in practice
The biggest limitation is availability. Block has not released Dessa publicly, and the current next step is security testing rather than commercial rollout. There are no announced prices, launch dates or customer integrations in the source material.
For now, the practical takeaway is to watch whether speech-native models become a new baseline for serious voice agents. If a system can interpret laughter, hesitation and interruptions without relying first on text conversion, it may handle messy real-world conversations better than a transcript-only pipeline.
But buyers should stay cautious. Until Dessa is available to test, its performance in everyday business settings remains an open question. The near-term signal is that major product companies are treating emotional and conversational nuance as a core problem for voice AI, not just a polish layer added after transcription.
Frequently asked questions
What is Dessa?
Dessa is an AI speech model unveiled by Block, the company behind Square and Cash App.
How was Dessa trained, according to Block?
Block says Dessa was trained entirely on human speech, with no text data, and was built from scratch.
Is Dessa publicly available?
No. Block says Dessa is not publicly available yet and is first seeking AI security researchers to test it for unexpected risks.
