You've probably felt that uncanny stiffness whenever you talk to a voice assistant—whether it's ChatGPT's standard voice mode, Siri, or a bank's phone support bot. The words are right, the grammar is spotless, the voice sounds pleasant enough, yet your brain immediately flags it: you're talking to a machine. The pauses between turns are unnaturally even, the tone doesn't shift when you crack a joke, and if you try to jump in mid-sentence, the bot either drones straight through its scripted paragraph or chokes and loses the thread.
The research team at Block (Jack Dorsey's fintech empire behind Square and Cash App) set out to kill that robotic stiffness for good. They unveiled an experimental voice system called Dessa that sounds so startlingly human that its creators decided to keep the weights locked away for now.
Death of the Classic Pipeline: Why Text Is No Longer Needed for Speech
To see why Dessa is such a fundamental break from the past, let's look under the hood of traditional voice AI. Historically, every conversational voice system ran on a clunky three-stage assembly line:
- Automatic Speech Recognition (ASR): your voice gets transcribed into plain text.
- Large Language Model (LLM): text tokens get processed to generate a written reply.
- Text-to-Speech (TTS): a voice synthesizer reads that text aloud.
See where the magic dies? The instant living human speech gets flattened into text, 90% of its emotional and contextual signal evaporates. A transcript can't tell whether you delivered a line deadpan or dripping with sarcasm, whether your voice trembled a bit, or whether you let out a heavy sigh or a chuckle before answering.
Dessa throws the text middleman straight into the bin. It's an end-to-end audio-to-audio neural network. It never converts speech into letters and back again. Trained directly on hundreds of thousands of hours of real human conversations, its internal layers operate on continuous spectral and acoustic patterns, understanding and generating speech purely through sound.
True Full Duplex: Interruptions, Laughter, and Emotional Awareness
Ditching the text bottleneck let Block's engineers build genuine full-duplex conversation. Standard voice bots run half-duplex: the mic is either listening to you or muted while the model talks.
Dessa listens to the audio stream continuously, even while speaking. If you start talking mid-sentence, it reacts like a real person sitting across from you at a coffee shop: instead of snapping silent like a tripped breaker, it smoothly lowers its volume, drops in a quick "mm-hmm" or "go ahead," and hands you the floor.
The micro-behaviors are even wilder:
- The model can laugh right in the middle of a phrase, catching its breath naturally just like a person does.
- It takes organic thinking pauses, drawing out vowels or adding slight hesitations when a question catches it off guard.
- If it hears frustration or exhaustion in your voice, Dessa immediately softens its timbre to a calmer, warmer tone without needing a single prompt instruction.
- Response latency—the gap between the end of your sentence and the start of hers—sits under 200 milliseconds, below the threshold where the human ear registers a pause as machine lag.
From a GB200 Supercluster to a Local MacBook: Inference Speed and Privacy
Training an end-to-end audio architecture like this took massive firepower. Block leaned on a supercluster powered by NVIDIA GB200 NVL72 chips to crack audio compression and continuous wave representations.
The biggest surprise for independent engineers, though, is how lean the model is at inference time. The team optimized the weights so aggressively that Dessa can run completely locally on a MacBook Pro with an Apple M4 Max chip.
Pause and think about what that does for privacy. You no longer have to ship intimate conversations, medical questions, or sensitive financial discussions off to remote corporate servers. A voice assistant with cinema-grade naturalness runs straight in your laptop's unified memory, completely offline.
So why hasn't Block open-sourced it? The answer is obvious: massive social engineering risks. Over a phone line, Dessa is indistinguishable from a live human being—a nightmare scenario for phone scammers and voice-biometric bank fraud. For now, the model remains restricted to a closed research preview for security specialists. Still, the direction is unmistakable: the era of text crutches in voice AI is over.

