Block Releases Dessa, a Speech-to-Speech AI Model Trained on Human Conversation

Block has released Dessa, an AI model trained from scratch on human conversation, according to the company. The model is designed to generate speech with emotion, nuance, pauses and a natural cadence.

Unlike conventional voice assistants that convert speech to text before responding, Dessa uses a speech-to-speech architecture to process and generate audio directly. Its duplex design allows the model to speak and listen simultaneously, avoiding the pauses typical of many AI assistants. Block posted several audio examples on social media, including a guessing game and a conversation featuring a different voice, to demonstrate the model’s realistic output.

The company said a small group of AI safety researchers has been given early access to Dessa, which can run locally on any MacBook Pro M4 Max or better. The Dessa team said the model was created out of curiosity and to improve understanding between humans and AI, as well as to address safety issues stemming from AI models being largely deaf to emotion in human voices.

Block, formerly known as Square, changed its name in 2021 and owns services including Cash App, Afterpay and Tidal.