Block has revealed a new voice model developed inside its research division that sounds remarkably human, incorporating natural pauses, laughs, and emotional inflections while operating entirely on local hardware. The financial technology company, led by Jack Dorsey, announced that the model, dubbed Dessa, was trained directly on human conversational speech using a single Nvidia GPU rack rather than relying on conventional text transcription pipelines.
The development represents a notable shift in how conversational artificial intelligence is constructed. By skipping the standard intermediate step of transcribing voice into text and back again, Dessa is designed to preserve the subtle auditory cues that define ordinary human conversation. Because the model operates full-duplex, it can listen and speak simultaneously, enabling real-time interruptions and spontaneous dialogue without awkward turn-taking delays.
Eliminating the Text Pipeline
Most existing voice assistants rely on a multi-stage architecture: an automatic speech recognition engine converts spoken audio into text, a large language model processes the text and drafts a response, and a text-to-speech synthesizer reads that response aloud. While functional, this pipeline discards critical non-verbal information, including pitch, cadence, hesitation, and emotional tone.
Other modern speech tools, such as ElevenLabs v4 or Fish Audio Drama 3, attempt to reintroduce emotion by applying explicit vocal tags to generated text. Block took a different path, building an end-to-end speech model where emotion and conversational nuances are embedded directly into the system.
In a blog post explaining the philosophy behind the model, the company was blunt about the shortcomings of current voice tools: "Interacting with current voice AI can just plain suck. It turns your voice into text and throws away an incredible amount of information text alone can’t carry. The most powerful AIs in the world can’t hear how you feel. We think that’s ridiculous, leads to a worse experience, and is potentially dangerous."
Block co-founder Jack Dorsey underscored that vision when sharing the project publicly. "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 said. "Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence."
Realistic Emotion and Full-Duplex Dialogue
Demonstrations shared by Block researchers highlight how radically different Dessa sounds compared to traditional voice interfaces. In published audio samples, the assistant adapts its vocal delivery to the context of the exchange. In one instance where the bot mentioned feeling tired after a day of work, its voice audibly dragged with fatigue. In another conversation, the voice sounded distinctly emotional and close to breaking down while describing feeling comfortable with a conversational partner.
The model also handles everyday, low-latency exchanges seamlessly. Demonstrations show users playing spontaneous guessing games with the system, where the assistant quickly interrupts, guesses, and reacts in real time.
Block researcher Ragavan Thurairatnam described the milestone as a pivotal breakthrough for audio AI, writing that the team had built "the first AI model you can have a realistic conversation with that was trained entirely on speech" and describing the release as a major inflection point for speech technology.
Key technical characteristics of Dessa include:
- Direct speech-to-speech architecture trained fully on human conversational recordings
- Full-duplex interaction, allowing simultaneous speaking, listening, and real-time interruption
- Native emotional cadence, including laughs, pauses, and mood shifts without secondary prompting tags
- Lightweight training footprint achieved on a single Nvidia GPU rack
- On-device inference capable of running locally on consumer machines
Running Locally on Consumer Hardware
One of the most consequential aspects of Dessa is its computational footprint. Despite producing expressive, full-duplex speech, Block says the model can execute locally on consumer-grade hardware, specifically pointing to laptops equipped with an Apple Silicon M4 Max processor or equivalent consumer GPUs.
Running conversational voice models on-device offers significant practical advantages:
- Latency reduction: Eliminating trips to remote cloud servers slashes conversational lag, which is essential for natural dialogue.
- Privacy protection: Voice interactions remain on the user's local hardware rather than streaming audio to external servers.
- Cost efficiency: Running models locally bypasses the steep ongoing server and bandwidth costs typically associated with cloud-hosted audio synthesis.
Safety Reviews and Future Applications
Block has not opened Dessa for general public access yet. Instead, the company has issued a call for external safety researchers to evaluate the model and test its boundaries. Given the uncanny realism of the synthesized voices and their ability to express persuasive emotion, voice agents of this caliber introduce serious risks around voice cloning, impersonation, and social engineering fraud. Rigorous red-teaming will be necessary before any broad consumer release.
Beyond research benchmarks, the commercial incentives for Block are substantial. As the parent company of Square, Block already integrates voice artificial intelligence into retail and restaurant environments, such as automated food ordering systems. Deploying a lightweight, highly expressive voice model directly on local point-of-sale hardware could drastically reduce infrastructure overhead while making voice-driven customer interactions feel far more natural.
If Block can prove that Dessa remains controllable and secure, this speech-native approach may accelerate a broader departure from text-tethered voice bots toward truly responsive, conversational computing.
Filed under: AI, TechNews, Software, VoiceAI, MachineLearning