Block, the company founded by former Twitter executive Jack Dorsey, has introduced a new voice AI model called Dessa. The main feature of the development is the complete rejection of text data during training: the system perceives and processes exclusively audio recordings of human conversations. According to the creators, this approach allows the model to capture the smallest nuances of spoken speech, including intonation, pauses, and emotional coloring, ensuring the most natural interaction possible.
Architecture Without a Text Layer
Traditional voice assistants and modern conversational AI models typically use a multi-stage architecture. First, the heard sound is converted into text using speech recognition modules, then the text request is processed by a language model, and finally, the response is synthesized back into a voice format. This approach inevitably leads to the loss of emotional shades, intonations, and the natural rhythm of the conversation.
Unlike classical systems, the Dessa model operates on direct audio signals without an intermediate text stage. This allows the artificial intelligence not only to hear the interlocutor's words but also to simultaneously analyze their emotional state. The system is capable of instantly reacting to sarcasm, laughter, voice fluctuations, or changes in tone right during the conversation, making communication virtually indistinguishable from a dialogue with a living person.
Full-Duplex Capabilities
The new development supports full-duplex communication, which means the ability to listen and speak simultaneously. Thanks to this, the artificial intelligence can pause in time, react to the interlocutor's remarks, or even interrupt them depending on the context of the conversation. During internal testing, the system demonstrated the ability to independently supplement forgotten movie titles or change its speech manner upon the user's request.
In addition to high adaptability, the model stands out for its hardware efficiency. To train it, engineers used only a single high-performance server with graphics processors, and the finished version is capable of operating locally on MacBook Pro laptops equipped with Apple M4 Max chips and higher. The audio processing speed exceeds real-time, completely eliminating delays in dialogue.
Release Prospects and Availability
At present, the Dessa voice model has not been released into open access for the general audience. The developers provide limited testing access exclusively to artificial intelligence safety specialists to conduct independent audits and checks. Other interested users can sign up for the waiting list on the company's official resources while the team completes preparations for a full-scale commercial launch.
