Block unveils Dessa, a localized voice AI with human inflection
Jack Dorsey's fintech firm introduces a full-duplex speech model capable of running on consumer hardware with native emotional expression.
Key highlights 路 3 min read
- Block, the financial services firm founded by Jack Dorsey, has introduced a new artificial intelligence research project called Dessa.
- The Scale Report understands that the model is designed to operate locally on consumer-grade hardware, specifically mentioning compatibility with systems like the MacBook M4 Max.
- Most current voice AI systems suffer from a loss of data because they convert voice input into text before processing it, stripping away the emotional context inherent in vocal delivery.
The Scale ReportBlock, the financial services firm founded by Jack Dorsey, has introduced a new artificial intelligence research project called Dessa. The model distinguishes itself from industry standards by functioning as a full-duplex, speech-to-speech system, enabling simultaneous talking and listening. Unlike competitors that rely on text-based pipelines to generate audio, Block claims Dessa was trained entirely on human speech to preserve nuanced markers like laughter, fatigue, and pauses.
Localized processing on consumer hardware
The Scale Report understands that the model is designed to operate locally on consumer-grade hardware, specifically mentioning compatibility with systems like the MacBook M4 Max. By bypassing the cloud-based text-to-speech architectures common in offerings from ElevenLabs or Fish Audio, Block suggests it has eliminated the artificial cadence often found in automated assistants. The company claims the model was developed using a single NVIDIA GPU rack.
Why it matters for voice AI
Most current voice AI systems suffer from a loss of data because they convert voice input into text before processing it, stripping away the emotional context inherent in vocal delivery. By integrating emotional capability directly into the speech model, Block is attempting to shift the industry away from the limitations of large language models that view voice as a secondary output layer. This approach mirrors the performance characteristics of OpenAI's GPT-Live.
Safety and commercial outlook
While Block is currently soliciting feedback from safety researchers to identify potential misuse or deception risks, the model remains unavailable for public access. The firm intends to address concerns regarding how hyper-realistic voice agents might be leveraged to impersonate or manipulate individuals. For a company like Block, which already integrates voice AI into Square for merchant services, an efficient on-device model could significantly reduce the costs of powering customer-facing commerce experiences.
Reporting based on coverage from AI News & Artificial Intelligence | TechCrunch.




