Block Debuts Dessa Voice AI Running Locally on M4 Max

Jack Dorsey’s payments company Block has unveiled Dessa, a voice AI model that processes speech directly without converting audio into text. In live demonstrations, the software listens while speaking, laughs mid-sentence, and pauses with natural human rhythm.

Block trained the model from scratch on a single Nvidia GB200 NVL72 computing system. The resulting weights are compact enough to run locally on an Apple MacBook Pro equipped with an M4 Max chip, keeping all voice data on the device.

The research preview has sparked intense debate. Observers see broad potential for children’s education, while cybersecurity experts warn that hyper-realistic voice tools could supercharge phone scams.

Technical Factsheet: Block Dessa Voice AI

How Dessa Skips the Text Bottleneck

Traditional voice tools like Siri, Alexa, or standard chatbot voice modes rely on three separate models chained together:

1. Speech to Text: An audio model transcribes spoken words into plain text tokens.
2. Text Generation: A large language model reads the text and writes a written reply.
3. Text to Speech: A voice generator converts that written text into synthetic audio.

This three-stage pipeline introduces noticeable lag. More importantly, it strips away critical audio information such as pitch, sighs, vocal hesitation, and emotional warmth.

Dessa bypasses text entirely. By training directly on human speech waveforms, the model processes audio natively. This allows full duplex conversation: Dessa listens while speaking and yields instantly when interrupted, just like a human conversation partner.

“Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence.”

Jack Dorsey, Head of Block

Education Hopes vs Scam Fears

Dessa’s natural conversational tone quickly drew contrasting reactions across the technology sector.

Binance founder Changpeng “CZ” Zhao pointed to potential breakthroughs in interactive learning for his non-profit education platform, GiggleAcademy:

“This might be very good for GiggleAcademy. Teaching kids in current ‘AI voice’ was a challenge.”

Changpeng Zhao, Founder of Binance & GiggleAcademy

However, the same humanlike realism raised urgent red flags regarding consumer fraud. Filmmaker Justine Bateman questioned the necessity of making AI sound human, arguing that it lowers people’s natural defenses against manipulation.

Journalist Shiv Aroor echoed those concerns, warning that telephone fraud could escalate dramatically: “Scam calls about to attain pandemic level.”

Why Local M4 Max Inference Matters

Most frontier voice models depend on massive cloud server farms. In contrast, Block claims Dessa runs locally on a laptop equipped with an Apple M4 Max processor.

Running on-device changes the operational equation in two ways:

• Complete Privacy: Raw audio never leaves the personal device, shielding sensitive voice conversations from cloud logs and corporate data collection.
• Zero Network Latency: Local execution eliminates the round-trip network delays of cloud APIs, making real-time voice interruptions immediate.

What Comes Next for Dessa

Block has not yet released the model’s parameter size, benchmark scores, or an official public release date. The demonstrations remain company-produced showcases rather than independently audited benchmarks.

For now, Block is limiting early access to an external cohort of AI safety researchers. The company notes that direct speech processing could help systems detect emotional distress and pain cues that text transcripts miss, while researchers test for misuse vectors before any broader rollout.