IT Home On October 25, Block, the company under Jack Dorsey, co-founder of Twitter, launched a new voice model called Dessa. It supports local operation and can simulate emotional expressions such as laughter, pauses, and sobs, giving a highly human-like sound. According to the official description, Dessa is fully trained on human voice data, and the training process used only a set of NVIDIA GPU racks.
Dessa belongs to the full-duplex model. Similar to GPT-Live, which OpenAI released earlier this year, Dessa can listen to user voices while speaking. Most mainstream voice models generally require a large language model to retrieve knowledge; Block said that Dessa’s own knowledge comes from the dialogue data used for its training.
According to Block, current voice AI experiences are often rigid and inconsistent. The root cause is that the generation process first converts them into text, and then the text drives voice synthesis. Many models such as ElevenLabs v4 and Fish Audio Drama 3 rely on voice labels to control emotions. However, this Dolsi company claims that emotion expression is natively built into the model. It remains unclear, though, whether it still relies on some kind of label for regulating emotions at the underlying level.
Block wrote on the blog: “The current voice AI interaction experience is really poor. It first converts your voice into text, and a large amount of information that cannot be conveyed in text is simply lost. Even the most powerful AI worldwide cannot detect your emotions. We think this is unreasonable; not only is the experience bad, but there are also potential risks.”
The research team has released multiple examples of conversations with Dessa. This AI can switch between different voices, speech rates, and emotions. For example, in one conversation, the AI says it is very tired at work, and its voice also reflects a sense of exhaustion.
In another example, the AI spoke about a feeling that brought comfort to the other person, with an emotional tone and almost a sob. Another demonstration shows Block’s employees playing games with this voice assistant.
Official description: This model can run locally on consumer-grade computers, with the requirement being a MacBook with M4 Max or higher specifications. However, at this stage, Dessa is not available for trial use. Block has sent an invitation to security researchers to conduct tests, focusing on scenarios where such voice AI agents may deceive humans.
According to IT Home, Block’s main business is financial services. It provides solutions for small and micro enterprises, and its Square platform already uses voice AI to handle order processing. After developing this on-device voice model, Block is expected to launch more voice products for merchants and ordinary users at lower costs.