Kimi K3.1 Pro: Moonshot’s open-weight bet on code + finance

What happened

Block released a research preview of a new speech model called DESA, developed by its in-house AI research lab. Unlike conventional systems that rely on text training, Block built DESA from scratch using raw human conversation audio. The model bypasses written captions and transcripts entirely, operating directly on spoken inputs and outputs without pre-trained baselines.

The architecture uses full duplex speech-to-speech processing, allowing the model to listen and speak simultaneously in real time. Rather than relying on a turn-taking interface, DESA is designed to register the natural cadence of speech, including pauses, interruptions, laughter, emphasis, and conversational uncertainty.

This approach cuts out the traditional three-stage pipeline where audio is converted to text, reasoned over, and synthesized back into speech. Training natively on spoken audio also helps the model capture real-world speech patterns, including unwritten dialects, slang, and half-sentences that standard text datasets typically miss.

Why it matters

Traditional voice assistants suffer from latency because translating speech to text and back creates unnatural pauses. In conversation, an extended delay breaks flow and trust. Eliminating text intermediaries transforms voice from a clunky command tool into a fluid collaboration medium, enabling automated agents to handle nuanced customer service tasks like scheduling, address changes, and refunds without rigid phone trees.

Strategically, native audio training challenges the assumption that advanced intelligence must stem from published writing. Text datasets overrepresent published authors, while raw audio captures how everyday people actually communicate across accents, dialects, and incomplete sentences. The operational advantage shifts toward natural interaction rather than prompt engineering.

However, lifelike voice tools elevate security and privacy concerns. Improved voice generation makes deepfake fraud easier, requiring strict verification for financial requests. Spoken audio also exposes personal emotion and background context, making clear data retention, storage, and deletion policies critical for enterprise deployment.

What to watch next

• Whether Block transitions DESA from an internal research preview into its commercial small business and customer products.
• Latency benchmarks comparing end-to-end speech models against traditional three-step speech-to-text-to-speech pipelines.
• Production performance on edge cases like overlapping speech, heavy background noise, and non-standard regional accents.
• Industry adoption of callback verification protocols to protect financial workflows against voice deepfakes.

What this means for you

Audit customer-facing communication channels for latency and operational friction. Identify routine tasks—such as address updates, appointment scheduling, and order inquiries—that can transition from rigid interactive voice menus to conversational voice agents. Ensure backend systems can retrieve data rapidly enough to match the speed of direct speech processing.

Implement strict verification protocols around voice communications. Treat spoken identity with the same caution as passwords by establishing mandatory callback procedures for money requests. Internally, practice voice-first prompting by instructing AI to ask clarifying questions before drafting, using back-and-forth dialogue to refine outputs rather than relying on one-way text commands.

Key takeaways

• Block introduced DESA, a research preview speech model trained natively on conversation audio without text transcripts.
• Full duplex speech-to-speech processing lets the system listen and speak simultaneously in real time.
• Training directly on audio captures natural timing, laughter, accents, and unwritten dialects while reducing pipeline latency.
• Realistic voice models heighten deepfake risks, requiring callback verification protocols for money requests.
• Directing AI through conversational back-and-forth dialogue yields better results than static, single-prompt instructions.

FAQ

What is Block's DESA speech model?

DESA is a speech-to-speech artificial intelligence model developed by Block’s in-house AI research lab as a research preview. Instead of learning from written text, DESA was trained from scratch directly on human conversation audio. It operates using a full duplex architecture that listens and speaks in real time without converting speech to text or using pre-trained models.

How does full duplex speech differ from standard voice AI?

Standard voice assistants operate like walkie-talkies, forcing users to take turns by speaking, waiting, and hoping the system understood. Full duplex speech enables the model to listen and speak at the same time in real time. This allows the system to register natural conversational timing, interruptions, pauses, uncertainty, and laughter.

Why did Block avoid text transcripts when training DESA?

Block avoided text transcripts to eliminate latency, improve naturalness, and expand language coverage. The traditional three-step pipeline of speech-to-text, reasoning, and text-to-speech introduces conversation-killing delays. Direct audio training captures vocal nuances like timing and emphasis while representing spoken dialects, slang, and half-sentences that are never written down in published text datasets.

What security risks do realistic voice models create?

Lifelike voice models make deepfake impersonation easier, raising fraud risks when bad actors request money over the phone. Because voice conveys personal identity and sensitive background details, users and enterprises must understand where audio is stored, whether it trains future models, and how to verify urgent financial requests via known numbers.