Block, the parent company of Square and Cash App, announced a speech AI model called Dessa. Block said it trained Dessa from scratch on recordings of human conversation, with no text and no pretrained models. Sound goes in and sound comes out, and the model can hear and speak at once. Block said it replies faster than real time and can run locally on a high-end MacBook Pro (M4 Max or better). It gave no training time or cost.
Block argues that turning speech into text loses stress, timing and tone. Stressing a different word in one sentence about stolen code changes the meaning each time. Block also says one spoken word takes thousands of times more bytes than text. It says Dessa learned language from hundreds of times fewer words than text-based models, but it gives no figures here, and LLMs also learn code and other material. These claims rest on Block's own demo clips; no outside test is cited. Block has set no release date or price, and offers early access only to a small group of AI safety researchers.
Teams that build or buy voice features can ask vendors for independent evaluation, latency figures, language coverage and safety findings. Until outside tests exist, Dessa's emotion handling, and its run on a high-end MacBook Pro (M4 Max or better), are Block's claims.