Voice AI technologies are moving towards developing more natural conversations, combining speed of response with simulation of human tone and rhythm of speech. In this context, Block, led by Jack Dorsey, has unveiled its voice model, "Dessa," which it says learns directly from human voice conversations, without relying on written texts for training.
Hot Topics
According to the model's official website, "Desa" aims to simulate aspects of human communication, such as laughter, pauses, changes in tone of voice, and speed of speech, making interaction with artificial intelligence closer to everyday conversations. However, these capabilities, which appear in experimental recordings, do not alone prove the model's superiority over its competitors, given the need for independent tests that measure its performance.
Some voice AI systems rely on converting the user’s speech into text, then processing the text and producing a response that is later converted into voice. Although this method is effective, it may miss some of the cues that accompany speech, such as hesitation, sarcasm, and subtle changes in tone.
Block says that Desa was trained from scratch using human voice conversations, without written scripts or pre-trained models. This approach allows the model to learn language patterns from the voice signals themselves, including word pronunciation, response timing, and the relationship between tone and meaning.
The company explains that the model can use laughter, pauses, and volume changes within a conversation, and it offers test recordings that include vocal expressions conveying different emotions. However, producing a voice that sounds affected or excited does not necessarily mean that the system understands human emotions as humans do; assessing this ability requires broader and independent testing.
Listening and speaking simultaneously
One of the most notable features of DESSA is its adoption of full-duplex audio technology, which allows the system to listen to the user and speak to him simultaneously, instead of waiting for one party to finish before the other begins.
This technology is useful in simulating the rhythm of human conversations, which sometimes include interruptions, brief signs of approval, and quick responses to what the other party is saying. According to the company, the model can determine the timing of its intervention in the conversation and use concise phrases that indicate follow-up, while minimizing delays in response.
Regarding the technical architecture, Block stated that she trained the model using an NVIDIA GB200 NVL72 computing system. She also indicated that it could be run locally on a MacBook Pro with an M4 Max processor or more powerful. Local operation may reduce reliance on remote servers, but its impact on privacy, cost, and performance depends on how applications are implemented and data is processed.
Potential uses and growing competition
These capabilities can support the development of voice assistants, customer service, educational applications, and interfaces that require continuous dialogue. Faster and more natural interaction may make using voice systems easier in some tasks, but it does not alone guarantee the accuracy of the answers or the quality of the services they provide.
Companies, including OpenAI and Eleven Labs, are competing to develop more realistic and responsive voice systems, and Block's difference lies in its reliance on direct voice training, rather than being limited to pathways that combine text processing and speech generation.
Tags:
technology