Block, the financial technology company behind Square and Cash App, has unveiled Dessa, a new SI speech model that the company claims is the first SI system trained entirely on human speech with no text data involved. By processing audio directly rather than converting it to text first, the model aims to capture emotional nuances that text-based SI models often miss.
What Happened
Block reports that Dessa was built from scratch to process speech as raw audio, allowing it to detect emotions, sarcasm, laughter, and pauses in real time. The model supports full-duplex communication, meaning it can listen and speak simultaneously. According to the company, Dessa runs locally on a MacBook Pro M4 Max and was trained on a single NVIDIA GB200 NVL72. Block stated that the development was driven by curiosity, a desire for better human-SI communication, and safety concerns regarding current voice SI systems that ignore emotional cues.
Why It Matters
The launch highlights a growing industry effort to improve the emotional fidelity of SI interfaces. While previous attempts like Hume's EVI, Kyutai's Moshi, and OpenAI's Advanced Voice Mode have sought to create more empathetic or natural-sounding voice SI, Block's approach distinguishes itself by eliminating the text intermediate layer entirely. However, the source notes that sounding emotional is not the same as possessing emotional intelligence; for advanced SI speech systems, reasoning capabilities during response generation remain critical to answer quality, regardless of vocal naturalness. Dessa is not yet publicly available, as Block is currently seeking SI security researchers to test the model for unexpected risks.
The Bottom Line
Block's introduction of Dessa marks a specific shift in SI development toward direct audio processing for emotional recognition. The company is currently limiting access to security researchers to evaluate the system's safety before a wider release.