Dessa, Block's voice AI that learned to converse without ever reading text

With Dessa, Block is experimenting with a voice model trained solely on human conversations, capable of listening and speaking simultaneously.

Can artificial intelligence learn to speak by listening, without relying on written language? Block, the company led by Jack Dorsey, is exploring that question with Dessa, a voice model trained entirely on recordings of human conversations. Unlike systems that transcribe speech into text before generating an answer, Dessa processes audio directly, listens while speaking, and adjusts its timing to the person on the other side of the conversation. Its demonstrations feature an AI that laughs mid-sentence, finishes an interrupted thought, and changes its intonation, all without relying on a pretrained language model. Learning language directly from speech

Voice assistants often combine several specialized systems. Speech recognition converts an utterance into text, a language model produces a response, and text-to-speech software turns that response back into audio. Newer systems can process audio natively, although text still plays an important role in how many of them are trained.

The team behind Dessa has taken a different approach. According to Block's project presentation, the model was trained from scratch on recorded conversations, without text data or pretrained models. It therefore had to learn the patterns of spoken language from acoustic signals, gradually developing associations between sounds, words, conversational behavior, and meaning.

The challenge extends beyond understanding sentences. Speech carries information that a written transcript cannot fully preserve: hesitation, breathing, emphasis, sarcasm, laughter, and emotional changes. The same words can communicate different intentions depending on how they are spoken. By working directly with audio, Dessa is designed to retain those characteristics during both comprehension and response generation.

Block says the model learned language from hundreds of times fewer words than conventional large language models. However, the company has not published the audio dataset's size, its composition, the methodology behind that comparison, or the models used as reference points. Listening and speaking at the same time

Dessa is a full-duplex speech model, meaning it can process incoming audio while generating speech. This allows conversations to move beyond the rigid exchange in which one person must finish speaking before the assistant can begin responding.

In recordings published by Block, Dessa offers brief acknowledgments while its conversation partner speaks, reacts to jokes, completes unfinished sentences, and adjusts to pauses. These short acknowledgments, commonly called backchannels, are familiar features of human conversation.

One demonstration shows the model laughing immediately when something amusing is said, rather than waiting for the speaker to finish. Another shows it raising its voice when asked to speak louder. In an anniversary-surprise conversation, Dessa responds with noticeable changes in breath, pitch, and emotional delivery.

Block describes the examples as live recordings without human steering of the model's responses. Subtitles were added afterward. Another demonstration places Dessa in conversation with GPT-Live-1, ChatGPT's voice model, illustrating interruptions and conversational timing between two AI systems.

The recordings offer examples of how the system behaves, but they are not a controlled evaluation of factual reliability, speech comprehension, or performance across extended conversations. Training on a single NVIDIA system

Dessa was built by a small research team within Block, led by Ragavan Thurairatnam. The group includes former members of Dessa, a Toronto-based deep learning company with a history in speech generation, and reports directly to Jack Dorsey.

The researchers say they trained the model using a single NVIDIA GB200 NVL72 system. This is a powerful rack-scale computing platform containing multiple accelerators, not a single consumer graphics card.

Inference is designed to be considerably more accessible. Block states that Dessa can run locally on a MacBook Pro with an M4 Max chip or better, generating audio faster than real time. This would allow the system to sustain a live conversation without relying on a remote server for every exchange.

However, Block has not disclosed the model's parameter count, training duration, dataset size, exact latency measurements, or public benchmark scores. No detailed technical paper has yet established its architecture or performance relative to competing audio models. Emotional understanding as an AI safety question

Beyond conversational fluency, Block presents emotional awareness as a motivation for developing Dessa. A voice assistant working only from a transcription may miss important signals when a person expresses pain, distress, or uncertainty through tone rather than words.

A native audio model could potentially respond to those signals without first reducing them to text. But generating emotionally convincing speech is different from accurately understanding another person's emotional state. A system that sounds empathetic does not necessarily recognize distress correctly or know how to respond appropriately.

Block has not published a clinical evaluation or a comparable safety study demonstrating that Dessa can reliably detect emotional distress or handle sensitive situations. The team acknowledges that training a conversational model directly on audio introduces new safety questions.

To investigate them, Block plans to provide early access to a small group of AI safety researchers. For now, Dessa remains a research demonstration. Public audio examples are available, but the company has not announced general access, downloadable model weights, or a documented public API.