Block says the model learned these conversational behaviors directly from recorded human speech. It can produce short acknowledgments, adjust its speaking volume and respond to changes in tone, including sarcasm, excitement, frustration and sadness. The system also generates vocal details such as laughter, sighs and variations in breathing that are often missing from AI-generated speech.
The company developed Dessa around the idea that spoken communication carries information that written transcripts cannot fully preserve. A person's tone, timing and emphasis can substantially change the meaning of a sentence, but those signals are frequently lost when voice assistants reduce audio to text.
Block argues that this limitation affects more than conversational quality. An AI assistant that cannot recognize distress or other emotional signals may respond inappropriately during sensitive situations. Dessa is intended to explore whether models trained directly on speech can better interpret those signals while maintaining a more natural conversational rhythm.
The research team trained Dessa entirely from scratch using human audio, without relying on text datasets or pretrained language models. According to Block, the model acquired language using hundreds of times fewer words than large language models typically require, despite having to learn additional characteristics of speech such as accents, inflection and conversational timing.
Training was completed using a single Nvidia GB200 NVL72 system. Block also says Dessa generates responses faster than real time, allowing conversations to proceed without noticeable delays. However, the company has not disclosed the model's parameter count, training dataset size, measured latency or standardized benchmark results.
Instead, Block published recordings intended to demonstrate how the system behaves during unscripted conversations. The company says the interactions were recorded live without directing the model's responses, with subtitles added afterward.
In one demonstration, Dessa responds to humor by laughing during a sentence rather than waiting for the speaker to finish. Another shows it increasing its volume when asked to speak louder. Additional recordings feature the model helping someone recall a movie title, reacting to an unexpected anniversary announcement and responding empathetically during a personal conversation.
The demonstrations also include a conversation between Dessa and GPT-Live-1, a separate AI voice model. Block clarified that the developer of GPT-Live-1 had no involvement in Dessa's development and had not evaluated or endorsed the project.
Other clips shared by the research team show different voices, speaking styles and emotional expressions. One demonstration features a guessing game in which Dessa correctly identifies objects but continues asking questions after being told it has already won, illustrating that realistic speech does not necessarily translate into consistently accurate conversational understanding.
Dessa remains a research preview rather than a commercially available product. Block has initially restricted access to a small number of AI safety researchers, who must submit applications using institutional email addresses associated with universities, companies or research organizations.
The company says outside testing is intended to identify potential problems before broader availability.
“A model that learned to talk from sound alone is new territory, and we want outside eyes on it,” said Ragavan Thurairatnam, who leads the Dessa research team.
A public waitlist is available for other interested users, although Block has not announced a general release date. The company plans to publish additional demonstrations on X.
Dessa was created by a small research group whose members previously worked at the Toronto-based deep learning startup of the same name, which Block acquired in 2020. The team has experience developing deep learning systems dating back to 2012 and has worked on generative speech technology for more than eight years.
Thurairatnam, a co-founder of the original Dessa startup, leads the group, which reports directly to Block co-founder and chairman Jack Dorsey. The project also lists researchers Louay Hazami, Hashiam Kadhim, Qi Liu, Rayhane Mama, Aydin Polat and Marc Tyndel.
The team is continuing to investigate several questions about speech-only AI, including how efficiently these systems acquire knowledge, whether they can learn additional languages and what kinds of reasoning or adaptation they can demonstrate during conversations.
For now, Dessa's capabilities are supported primarily by Block's demonstrations and technical claims rather than independently published performance evaluations. The recordings show how a model trained without text can produce expressive, overlapping speech, but broader testing will be needed to establish how reliably those behaviors work across different users and situations.
Block has not announced plans to integrate Dessa into its existing products. The research preview instead provides an early look at an alternative approach to conversational AI, one that treats the sound and emotional characteristics of human speech as central to the interaction rather than information to discard during transcription.
This analysis is based on reporting from Gizmodo & Block.
Image courtesy of the-decoder.
This article was generated with AI assistance and reviewed for accuracy and quality.
