14:43Machine-translated from Azerbaijani by AIRead in Azerbaijani
Block Unveils Dessa Voice AI Model That Listens and Speaks at the Same Time
Jack Dorsey's company Block has unveiled Dessa, an artificial intelligence model designed for more natural voice interactions with people. According to the company, the model was trained from scratch using only human speech, without processing written data. Dessa can convey emotions, pauses, speech rhythm, and subtle vocal nuances. Square changed its name to Block in 2021.
Dessa directly processes the audio it hears and responds by voice. Conventional voice models first convert speech into text, then prepare a response and vocalize it. Dessa, however, can speak and listen at the same time. This architecture should reduce abrupt conversational stops and artificial pauses when the user interrupts. Ragavan Thurairatnam, head of the Dessa team, shared several audio and video samples of the model conversing with people. In a guessing game, the model correctly identified the everyday object the user was thinking of on its first try 2 times in a row. However, it continued guessing even after being told it had won. Despite this error, its tone of voice, pauses, and reactions sounded very close to human speech. In another conversation, Dessa initially mistook the user for someone named Ronaldo. While discussing Christopher Nolan's films, it also unexpectedly switched to Superman films directed by Bryan Singer. In another test, a user talks about a surprise trip to Paris, and Dessa responds in an emotional and grateful voice. In another audio sample, Dessa says in a style resembling the speech pattern of an African American man that it is tired, is going to work, and is picking products. During the conversation, it also talks about its newborn son and says it named him Darius Johnson. In side-by-side comparisons, Dessa sounded more fluent and natural than ChatGPT and GPT-Live-1, a model developed by OpenAI. The model can run on a MacBook Pro M4 Max or a more powerful computer without connecting to an internet server. Block has already granted early access to a small group researching artificial intelligence safety. According to the team, the project was developed to promote curiosity, better understanding between humans and artificial intelligence, and safety. The developers believe that current voice models have difficulty detecting human pain and emotions from a person's voice.
Dessa directly processes the audio it hears and responds by voice. Conventional voice models first convert speech into text, then prepare a response and vocalize it. Dessa, however, can speak and listen at the same time. This architecture should reduce abrupt conversational stops and artificial pauses when the user interrupts. Ragavan Thurairatnam, head of the Dessa team, shared several audio and video samples of the model conversing with people. In a guessing game, the model correctly identified the everyday object the user was thinking of on its first try 2 times in a row. However, it continued guessing even after being told it had won. Despite this error, its tone of voice, pauses, and reactions sounded very close to human speech. In another conversation, Dessa initially mistook the user for someone named Ronaldo. While discussing Christopher Nolan's films, it also unexpectedly switched to Superman films directed by Bryan Singer. In another test, a user talks about a surprise trip to Paris, and Dessa responds in an emotional and grateful voice. In another audio sample, Dessa says in a style resembling the speech pattern of an African American man that it is tired, is going to work, and is picking products. During the conversation, it also talks about its newborn son and says it named him Darius Johnson. In side-by-side comparisons, Dessa sounded more fluent and natural than ChatGPT and GPT-Live-1, a model developed by OpenAI. The model can run on a MacBook Pro M4 Max or a more powerful computer without connecting to an internet server. Block has already granted early access to a small group researching artificial intelligence safety. According to the team, the project was developed to promote curiosity, better understanding between humans and artificial intelligence, and safety. The developers believe that current voice models have difficulty detecting human pain and emotions from a person's voice.