For over a decade, consumer voice interfaces—from Apple’s Siri to OpenAI’s Advanced Voice Mode—have relied on an inherent architectural compromise: speech is captured as sound, transcribed into text tokens, processed through a linguistic reasoning engine, and synthesized back into spoken phonemes. While effective, the multi-stage pipeline introduces latency bottlenecks and strips away the rich, non-verbal vocal textures that define authentic human dialogue.

Now, an advanced research division within Jack Dorsey’s fintech conglomerate Block has broken fundamentally with that paradigm, unveiling an end-to-end audio speech model christened Dessa.

Reviving Dessa: From 2020 Acquisition to Native Speech Breakthrough

The initiative is directed by Ragavan Thurairatnam, head of Block’s artificial intelligence research lab. The debut represents a significant full-circle milestone: Thurairatnam originally co-founded the Toronto-based machine learning enterprise Dessa, which Dorsey’s company (then Square, Inc.) acquired in February 2020. While Dessa initially built foundational enterprise deep learning systems, Thurairatnam’s team has now repurposed the moniker for Block’s frontier speech-to-speech architecture.

Officially presented on October 23, 2026 as a Research Preview on block.xyz/dessa, the project departs radically from conventional generative speech systems by dispensing with written text altogether.

Abandoning Text: How Dessa Learns Like a Child

Engineers at Block built Dessa without using a single word of written text. The model incorporates no tokenizers, no dictionary databases, and no pre-trained foundational text models like GPT-4 or Claude. Instead, Dessa was exposed directly to tens of thousands of hours of raw acoustic recordings of genuine, unstructured conversations between human beings.

In technical documentation shared on Block’s portal, the model’s linguistic acquisition is compared directly to early human developmental cognition: an infant does not read an encyclopedia to learn vocabulary; it listens, internalizes tonal contours, discovers phonemic cadence, and gradually speaks back.

Because sound is processed natively across frequency spectrographs, Dessa retains the subtle acoustic markers that text conversion routinely flattens:

1. Vocal Inflection & Irony: Sarcastic tone, dry humor, and empathetic warmth are captured naturally through continuous pitch modulation.
2. Conversational Breathing: Unlike synthesized text-to-speech that produces unnaturally uniform breathing breaks, Dessa hesitates, takes realistic breaths, and articulates throat sounds organic to spontaneous dialogue.
3. Mid-Sentence Emotion & Backchanneling: Laughter, vocal chuckles, and thoughtful murmurs (“yeah,” “hmm,” “got it”) blend directly into spoken sentences without triggering jarring acoustic shifts.

Full-Duplex Architecture and Offline Apple Silicon Execution

Perhaps the most startling operational breakthrough demonstrated in Block’s announcement is Dessa’s full-duplex bidirectional stream. Traditional AI voice models are fundamentally turn-based: a user speaks, waits for the server to process, listens to the response, and cannot interject without abruptly cutting the model’s playback track.

Dessa listens and generates audio simultaneously on a continuous acoustic loop. If a user interrupts mid-sentence, the model does not freeze or reset; it adjusts its cadence, yields the conversational floor, or interjects naturally, replicating the fluid cadence of human telephone calls.

Equally consequential for privacy advocates and enterprise developers, Block verified that Dessa runs entirely locally and offline on consumer-grade hardware—specifically operating faster than real time on an Apple MacBook Pro equipped with an M4 Max chip. By stripping out bloated parameter sets associated with multilingual text corpora, the lightweight acoustic model executes low-latency inference without requiring round-trip cloud API handshakes.

Research Preview and Safety Waitlist

Rather than immediately deploying Dessa across Block’s consumer platforms like Cash App or Square Seller terminals, Block has launched the system strictly as an experimental Research Preview. Visitors to block.xyz/dessa can audit authentic unedited audio interactions demonstrating real-time interruptions, laughter, and dialect nuance.

To confront safety imperatives, Block has restricted initial hands-on access to a curated waitlist of speech scientists and AI safety researchers. Their mandate is to stress-test the model’s conversational boundaries, evaluate voice-cloning vulnerabilities, and establish verifiable provenance safeguards before any broader commercial deployment. Nonetheless, Dorsey and Thurairatnam’s demonstration makes one technological trend undeniable: the era of translating speech to text before thinking is rapidly coming to an end.