Jack Dorsey's Block has unveiled a voice model that can express laughter, pauses, and emotional tones with what the company describes as strangely human quality, according to an announcement from the company's research lab on October 23, 2026. The model, named Dessa, was developed using only human speech recordings and trained on a single NVIDIA GPU rack. Block positions the release as a breakthrough in conversational AI that eliminates the robotic quality plaguing existing voice assistants.
Dessa operates as a full-duplex model, enabling it to simultaneously speak and listen in the same manner as OpenAI's GPT-Live model introduced earlier in 2026, the company said. Unlike conventional speech systems that pair with large language models for information retrieval, Dessa draws its knowledge directly from conversational data used in training. The model can operate locally on consumer-grade hardware including a MacBook M4 Max or superior machines. Block's researchers demonstrated multiple conversations showing the agent adopting different vocal tones, pacing, and emotional expressions—in one example, the bot claimed to feel tired from work and sounded fatigued accordingly, while another clip captured the assistant sounding emotional and nearly breaking down while discussing comfort.
Block argued that existing voice AI solutions "can just plain suck" because they convert speech to text, discarding vast amounts of information that text alone cannot capture, according to a company blog post. The firm claimed that emotion is integrated into Dessa's architecture, contrasting with models like ElevenLabs v4 or Fish Audio Drama 3 that rely on voice tags to generate emotional speech—though Block did not clarify whether any tagging mechanism steers Dessa's emotional output. "We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built," Dorsey wrote in his announcement.
Current voice models produce unnatural and irritating results because text sits in the generation pipeline, Block explained, leading to what the company characterized as "a worse experience" that is "potentially dangerous" since powerful AI systems cannot detect how users feel. By training exclusively on speech rather than text intermediaries, Dessa preserves the nuance, cadence, and emotional content that traditional architectures strip away. Block's business—focused on finance and small business solutions—already uses voice AI for order handling through its Square product, but an on-device model promises to enable substantially more business and consumer applications at reduced expense. The company has opened applications for safety researchers to test Dessa before any public release, likely to evaluate scenarios where a realistic voice agent could deceive listeners. For enterprises weighing voice interface strategies, the arrival of speech-native models may reshape expectations around what conversational AI can deliver. Whether emotional authenticity proves essential or unsettling will depend heavily on how users respond when the technology escapes the lab.