Block releases Dessa voice model capable of local execution on laptops

Jack Dorsey’s financial technology company Block has released a new artificial intelligence voice model named Dessa from its research division, TechCrunch AI reported. The model incorporates natural human expressions including pauses, laughs, and cries during speech generation. [1]

TechCrunch AI reported that Dessa operates as a full-duplex model, allowing it to process incoming audio while simultaneously generating speech output. Block stated that the system derives its knowledge base directly from conversational training data rather than relying on a separate large language model for information retrieval. [1]

Block stated that Dessa was trained using a single Nvidia graphics processing unit rack and is capable of running locally on an Apple MacBook M4 Max computer, according to TechCrunch AI. As FTMQ SI reported earlier, Ragavan Thurairatnam serves as the project lead for Dessa at Block, where the model was trained from scratch exclusively on human speech interactions. [1][3]

Full-duplex audio capabilities mirror systems such as OpenAI's GPT-Live model, TechCrunch AI reported. OpenAI originally launched its ChatGPT platform in November 2022 using generative pre-trained transformers to process multimodal user prompts. [1][2]