Jack Dorsey's company Block has released a new voice model called Dessa. The company says it can laugh, pause and cry, and that talking to it feels like talking to a person. It runs locally on a single rack of NVIDIA chips.
Dessa can listen and speak at the same time. That is the same idea behind OpenAI's GPT-Live model. Block says the feeling in the voice is built in, not added later with text tags.
π The Backstory
Most voice AI tools turn speech into text, let a language model write an answer, then turn the text back into sound. That steps throw away tone, timing and feeling. Block argues this is why many voice assistants sound flat.
Block trained Dessa only on recorded human speech. The company posted sample talks with different voices and moods, but has not said when the public can use it.
π― Why It Matters
Better voice AI could make talking to computers feel natural. It also makes fake voices harder to spot.
Jack Dorsey's company Block has released a new voice model called Dessa. The company says it can laugh, pause and cry, and that talking to it feels like talking to a person. It runs locally on a single rack of NVIDIA chips.
Dessa can listen and speak at the same time. That is the same idea behind OpenAI's GPT-Live model. Block says the feeling in the voice is built in, not added later with text tags.
Most voice AI tools turn speech into text, let a language model write an answer, then turn the text back into sound. That steps throw away tone, timing and feeling. Block argues this is why many voice assistants sound flat.
Block trained Dessa only on recorded human speech. The company posted sample talks with different voices and moods, but has not said when the public can use it.
Better voice AI could make talking to computers feel natural. It also makes fake voices harder to spot.