Block Launches Dessa AI Model for Natural Speech Interaction

This innovation could redefine human-AI communication, impacting industries reliant on voice interfaces.

• Block announced Dessa, a full-duplex speech-to-speech AI model trained exclusively on human conversation audio.
• Dessa enables real-time conversation, processing audio input directly to audio output without text transcription.
• Early access is limited to select AI safety researchers, with a public waitlist for broader interest.

• Block acquired the Toronto-based startup Dessa in 2020, leveraging its expertise in generative speech models.
• The model aims to address limitations in conventional voice systems that lose emotional and tonal nuances through text intermediaries.
• Dessa runs locally on high-performance hardware, emphasizing privacy and real-time processing capabilities.

On October 23, 2026, Block unveiled Dessa, a groundbreaking AI voice model that processes audio directly, bypassing traditional text transcription methods. This innovation is significant because it allows for full-duplex communication, meaning that conversations can flow naturally with interruptions and emotional nuances intact. The model was developed by a specialized team within Block's AI research lab, which reports directly to co-founder Jack Dorsey. This structure emphasizes the importance of the project within Block's broader strategy to enhance human-AI interactions.

The decision to train Dessa exclusively on human conversation audio stems from a critical observation: conventional voice systems often strip away emotional and tonal information when converting speech to text. This loss can create safety gaps as AI systems gain more autonomy and are deployed in sensitive contexts. By focusing on audio-only training, Dessa aims to maintain the richness of human conversation, potentially leading to more empathetic and context-aware AI interactions.

The model's architecture is designed to run locally on advanced hardware, specifically the MacBook Pro M4 Max or better, after being trained on a single NVIDIA GB200 NVL72 system. This local processing capability not only enhances privacy but also reduces latency, allowing for real-time conversations that feel more organic. The initial release is positioned as a research preview, inviting AI safety researchers to explore its emotional awareness and conversational behaviors. This cautious approach reflects a commitment to understanding the implications of such technology before widespread deployment.

The announcement generated significant interest in technology circles, with many outlets highlighting the innovative training approach and the potential for improved human-AI emotional understanding. However, it is important to note that no model weights, API, pricing, or benchmarks were released, indicating that Block is still in the exploratory phase of this project. The focus remains on curiosity-driven exploration and AI safety evaluation, rather than immediate commercial applications.

• AI safety researchers: They will have early access to evaluate the model's emotional awareness and conversational behaviors.
• Tech companies: Organizations relying on voice interfaces may need to adapt to new standards in AI communication.
• Consumers: Users of AI-driven voice assistants could experience more natural interactions as this technology matures.

• Public waitlist engagement: Monitoring the number of sign-ups could indicate market interest and potential user adoption.
• Feedback from AI safety researchers: Insights from early evaluations will shape future iterations and safety protocols for Dessa.
• Competitor responses: Other tech companies may accelerate their own AI voice model developments in reaction to Dessa's capabilities.