Block, the digital payments and financial services company created by Jack Dorsey and formerly known as Square, has introduced Dessa, its own artificial intelligence tool. It is a voice model that has been trained from scratch using exclusively real conversations, so it is capable of providing a very human-like interaction, including expressions such as laughter, pauses or crying.
As reported by the company, Dessa is a full-duplex model, which means it can talk and listen simultaneously, as is the case with GPT-Live, launched by OpenAI earlier this year. They claim that it has been created using only sound and has learned to maintain conversations in real time just by listening; and in this sense, Block maintains that “It is the first AI in the world with which it is possible to have a realistic live conversation”.
This differentiates it from other systems that generally require a large-scale language model for information retrieval and pre-text processing to generate speech, which can make them appear unnatural or even annoying. “We believe that voice is the most natural way to interact with an intelligence; That’s why we created Dessa, the most authentic *full-duplex* voice model that has been developed to date.”; they have stated from the company. “Talking to him is like talking to a person: it is full of emotions, nuances, pauses and a natural cadence”.
To illustrate this, Block has shared several examples of conversations between team members and Dessa, as have Jack Dorsey himself or Ragavan Thurairatnam, Co-Founder of Dessa, through social networks. In the videos you can see how the tool is capable of starting conversations with different voices, cadences and emotional states. In one of the examples, the human interlocutor tells the AI that he has bought tickets for a trip to Paris, and the AI appears deeply surprised and emotional, almost close to crying.
The examples shared show that the tool decides what to say, when to say it and how to do it, providing its own voice and personality, depending on the conversation. Thus, he is capable of laughing at the exact moment when something is funny, even in the middle of the human interlocutor’s sentence; or to increase the volume if asked to speak louder. You can also complete the sentence when the interlocutor becomes abstracted; provide a wide variety of intonations and speech rates; and when to issue supportive responses, such as “yes!”, “I understand,” or “oh,” to show that you are following the conversation.
In one of the shared videos, Block presents a conversation between ChatGPT-Live-1 and Dessa to highlight the differences between both models.
Block defends its commitment to a speech-to-speech model due to the superiority of its responses compared to tools built on text. He believes that, under this approach, the models achieve a better understanding between humans and AI, and regrets that most current voice tools do not take advantage of that potential.
“As AI acquires greater capacity and independence, it remains practically deaf to the emotions we transmit through our voice“, they say from Block. “Imagine that you are suffering from intense pain and the voice AI, not perceiving it, continues to function as if nothing had happened. That is the problem related to security. Dessa is a step towards your solution”.
Precisely to guarantee security, the company has announced that it will give early access to a small group of researchers specialized in artificial intelligence security, in order to identify risks that have not yet been foreseen and perfect the system.
More info.: Dessa
