started · updated
Block unveils Dessa, a speech-to-speech AI trained only on human audio
Block, the fintech company led by Jack Dorsey, has announced Dessa, a research preview of a new voice AI model. Unlike traditional voice assistants that convert speech to text before processing, Dessa is a full-duplex, speech-to-speech model trained exclusively on raw human conversation without the use of text or pre-trained language models.
Because it bypasses the text conversion step, Dessa is designed to capture emotional nuances that text often loses, such as sarcasm, laughter, sighs, and changes in tone. The model can listen and speak simultaneously, allowing it to be interrupted or to interject with natural fillers like “yeah” or “got it.”
Technical demonstrations show the AI exhibiting human-like behaviors, including laughing mid-sentence or sounding exhausted. Block reported that the model was trained on a single NVIDIA GB200 NVL72 rack and is efficient enough to run locally on consumer hardware, specifically a MacBook Pro with an M4 Max chip or better.
Currently, Dessa is not available to the public. Block is providing early access to a small group of AI safety researchers to identify potential risks, such as the potential for highly realistic voices to be used in fraudulent activities.
Entities
Apple · Block · ChatGPT · Dessa · Jack Dorsey · Nvidia · OpenAI
Claims
What the coverage asserts, and how many sources carry each claim.
- [● 2 SOURCES] Critics have expressed concerns that the realistic voice could be used for scams and deception. www.pymnts.com · www.smartphonology.it
- [● 4 SOURCES] The model was trained using a single NVIDIA GB200 NVL72 rack. www.smartphonology.it · www.itmedia.co.jp · www.bug.hr · pc.watch.impress.co.jp
- [● 3 SOURCES] The technology is currently unavailable for general public use. www.news9live.com · www.bug.hr · pc.watch.impress.co.jp
- [● 8 SOURCES] Dessa uses full-duplex technology to listen and speak at the same time. www.smartphonology.it · www.news9live.com · www.itmedia.co.jp · www.bug.hr · shiftdelete.net · +3 more
- [● 8 SOURCES] Block introduced a voice AI model named Dessa. www.smartphonology.it · www.news9live.com · www.pymnts.com · www.itmedia.co.jp · www.bug.hr · +3 more
- [● 5 SOURCES] Dessa can run locally on a MacBook Pro with an M4 Max chip or higher. www.smartphonology.it · www.itmedia.co.jp · www.bug.hr · www.teknoblog.com · pc.watch.impress.co.jp
- [● 8 SOURCES] The model was trained exclusively on human conversation recordings without using text or pre-trained language models. www.smartphonology.it · www.news9live.com · www.pymnts.com · www.itmedia.co.jp · www.bug.hr · +3 more
- [● 6 SOURCES] Dessa can recognize and reproduce non-verbal cues like laughter, sarcasm, and sighs. www.itmedia.co.jp · www.smartphonology.it · www.bug.hr · www.news9live.com · www.teknoblog.com · +1 more