Block's Dessa voice model achieves full-duplex speech with local GPU inference
Block releases Dessa, a full-duplex speech-to-speech model that runs on consumer GPUs and generates emotion natively without text-intermediate steps.
Last verified:
Block, the fintech company led by Jack Dorsey, released Dessa, a full-duplex voice model that generates emotionally-responsive speech without intermediate text conversion and runs on consumer-grade GPUs. According to TechCrunch AI, the model is trained entirely on speech data and can operate locally on hardware such as a MacBook M4 Max, marking a departure from cloud-dependent voice systems that rely on text pipelines.
Dessa’s speech-to-speech architecture
Block’s approach differs fundamentally from current voice AI production pipelines. According to TechCrunch AI, models like ElevenLabs v4 and Fish Audio Drama 3 use voice tags to steer emotional tone during generation—a process that Block researchers argue introduces artificiality. Dessa instead embeds emotion directly into the learned representations, eliminating the text-intermediate stage that Block contends discards vocal information critical to human communication.
The company published demonstration conversations showing the agent expressing tiredness and emotional vulnerability with corresponding vocal changes—shifts in cadence, breath patterns, and prosody that emerge from the model itself rather than external control signals. TechCrunch AI reports that these examples were recorded from Dessa’s outputs, though Block has not published independent reproducibility benchmarks or disclosed the size of the training corpus.
Full-duplex capability and comparison to GPT-Live
Dessa’s ability to listen and speak simultaneously places it in a small category alongside OpenAI’s GPT-Live, released earlier in 2026. Full-duplex voice interactions eliminate turn-taking delays, creating a conversational experience closer to human-to-human dialogue. Block stated that the model draws knowledge from conversational context rather than relying on a separately-paired language model for retrieval, though Block has not disclosed how the model handles factual grounding or hallucination.
Deployment constraints and safety testing
Block is not yet releasing Dessa to the public. According to TechCrunch AI, the company is actively recruiting safety researchers to test scenarios where voice agents might deceive or manipulate interlocutors—a recognition that emotionally-responsive speech models pose unique social-engineering risks. Running inference on consumer hardware eliminates cloud infrastructure costs and latency, but it also makes prompt-injection and jailbreak attempts harder for Block to monitor at scale.
Business integration path
Block’s core products—Square payment processing and cash-management services—already incorporate voice-based ordering systems. According to TechCrunch AI, a proprietary speech-to-speech model running locally could reduce computational costs per interaction and unlock new consumer and small-business features. Block did not announce specific deployment timelines or pricing.
Why This Matters
Dessa signals that speech-native machine learning—training directly on audio rather than text representations—is becoming viable at scale. If Block’s claims about emotion encoding hold up under independent scrutiny, this architecture could shift voice AI from a text-generation derivative to a first-class modality. This matters for fintech companies optimizing transaction friction (where voice ordering gains speed over manual entry) and for any product where emotional authenticity in agent-to-human interaction affects trust and engagement. However, the absence of public benchmarks, the lack of confirmed training-set scale, and the early-stage safety research suggest the technology’s real-world reliability remains unproven outside Block’s controlled demonstrations.
Frequently Asked Questions
What makes Dessa different from existing voice AI models like ElevenLabs?
Dessa is trained entirely on speech rather than converting voice to text as an intermediate step. This allows emotion and vocal nuance to be embedded directly in the model rather than controlled via voice tags, according to Block's claims.
Can I use Dessa today?
Not yet. Block is currently limiting access to safety researchers to test edge cases where voice agents might manipulate users. Public availability has not been announced.
What is full-duplex speech?
Full-duplex means the model can both listen and speak simultaneously, unlike turn-based models. OpenAI's GPT-Live uses the same capability.