Explore Block's end-to-end voice model, built from scratch without text or pretrained language models. Listen to the official two-track recordings, check whether your Mac can run it locally, and try a live, interruptible voice AI right in your browser.
Zero-text training. Learned only from human conversation — no text, no pretrained models — on a single NVIDIA GB200 NVL72.
Full-duplex flow. Listens while it talks: laughs mid-sentence and drops in "yeah" or "got it" without waiting its turn.
Runs locally. Block says it runs faster than real time on a MacBook Pro with M4 Max or better.
Free: official recordings breakdown, M4 Max checker and BYOK mode
Independent site · Not affiliated with Block, Inc.
Live captions of what you say will appear here.
Live engine: real-time interruptible voice AI, not Block's Dessa weights.
Official Dessa recordings
Six live recordings published by Block’s Dessa research team on block.xyz/dessa. Each card plays a short excerpt of the two synchronized speaker tracks: the person on top, Dessa AI below. Start one and any other card stops.
It laughs the moment something’s funny, mid-sentence, the way people do. Other models wait for you to finish, then laugh on cue. Ask it to speak up and it gets louder. Nobody taught it that.
The person blanks on a movie title. Dessa finishes the sentence. It also knows a surprising amount, considering every fact it has come from raw sound.
A surprise. Dessa gasps “oh my god” under its breath, then answers with what sounds like teary gratitude.
Huge range in intonation and speed.
Here it sounds genuinely empathetic.
And here it talks to ChatGPT Live.
Block’s disclaimer: In this demo, Dessa is talking with GPT-Live-1, a third-party AI model. Both voices are AI-generated. The developer of GPT-Live-1 is not affiliated with Block and has not reviewed or endorsed Dessa.
Short excerpts are shown for commentary and discussion. All recordings, titles, and captions belong to Block, Inc. DessaVoice is an independent site and is not affiliated with Block.
How Dessa works
Most voice assistants turn your speech into text, think in text, then read text back aloud. Block's Dessa skips text entirely: audio in, audio out.
Text in the middle
Zero text
Dessa claims are from Block's official Dessa page. DessaVoice is an independent site and is not affiliated with Block, Inc.
Local readiness
Block says Dessa's inference runs faster than real time and runs locally on a MacBook Pro with M4 Max or better. The model weights are not publicly released yet, so any result below is an estimate, not a test of your machine.
Unified memory on Apple silicon; system RAM on a PC.
Estimated result
Matches Block's stated targetThis is M4 Max-class Apple silicon with ample memory, which matches the hardware Block says Dessa runs on locally. Actual requirements will only be known once weights are released.
Estimate based on Block's public statement; not a benchmark. Not affiliated with Block, Inc.
Block has not published model size, memory requirements, or release dates for Dessa weights. Our live voice calls do not run Block's Dessa model.
We'll email you if Block publishes downloadable weights or community MLX/GGUF builds become available.
Compare voice models
How Block's Dessa compares with other conversational voice models, based only on what each lab has published. “Not disclosed” means the lab has not documented it publicly.
| Feature | Block Dessa | Sesame AI (CSM) | OpenAI GPT-Live-1 | Kyutai Moshi |
|---|---|---|---|---|
| Core architecture | End-to-end speech-to-speech: audio in, audio out, no text[1] | Llama backbone plus audio decoder; speech from text and context[2][3] | Full-duplex voice model; delegates reasoning to a backend model[4][5] | Speech-text model: 7B Temporal Transformer, Depth Transformer, Mimi codec[8][9] |
| Training data | Only human conversation audio; no text, no pretrained models[1] | About 1M hours of transcribed public audio, mostly English[2] | Not disclosed | Helium text LLM init; 7M audio hours; Fisher; synthetic dialogues[8] |
| Full-duplex & backchanneling | Yes: listens while talking, cuts in and backchannels[1] | Not in CSM; research post lists full duplex as future work[2] | Yes: listens while speaking, handles interruptions and backchannels[5][6][7] | Yes: parallel user and Moshi audio streams, overlap and interjections[8][9] |
| Local / open weights | Runs locally on MacBook Pro M4 Max or better; weights not released[1] | Base CSM-1B open weights, Apache 2.0, on Hugging Face[3] | Hosted API only; no public weights[4] | Open weights (CC-BY 4.0); MLX version runs on Mac[9] |
| Availability | Research preview; waitlist at block.xyz/dessa[1] | Research preview app; CSM-1B code on GitHub[2][3] | OpenAI API since September 10, 2026; $0.05 per minute[4][6][7] | Open source on GitHub and Hugging Face; demo at moshi.chat[9] |
FAQ
Straight answers about Block's Dessa model and what you can do on DessaVoice today.
Dessa is a full-duplex speech model from a small AI research lab at Block, released as a research preview. Block says it was trained from scratch on nothing but human conversation, with no text and no pretrained models. It takes audio in and gives audio out, listens and talks at the same time, and knows when to cut in or backchannel with a “yeah!” or “got it!”.
Block's official Dessa model is a research preview that is only available through a waitlist at block.xyz/dessa. On DessaVoice, the breakdown of Block's official recordings, the MacBook hardware readiness checker and bring-your-own-key (BYOK) mode with your own OpenAI key are free. Live hosted voice calls require the $4.90 Instant Voice Pass. Our live engine is a real-time interruptible voice AI, not Block's Dessa model.
No. Both dessa-ai.com and DessaVoice (dessavoiceai.org) are independent websites that are not affiliated with Block, Inc. The official Dessa page is block.xyz/dessa.
Block says Dessa's inference runs faster than real time and that it runs locally on a MacBook Pro M4 Max or better. However, Block has not publicly released Dessa's model weights as of now, so you cannot download and run it yet. Our free hardware checker shows how your Mac compares with that M4 Max baseline.
Block says Dessa was trained only on human conversation audio, with no text and no pretrained models, and that it is full-duplex with natural backchannels. Sesame's published Conversational Speech Model (CSM) is a text-and-speech model with a Llama backbone, trained on about one million hours of transcribed public audio, and Sesame's research post lists full-duplex modeling as future work. Sesame has open-sourced a 1B CSM checkpoint under Apache 2.0, while Dessa is a waitlist-only research preview.
The model shares its name with Dessa, the Toronto deep learning startup that Square (now Block) acquired in 2020. The Block research lab behind the model was founded by former members of that startup.