Voice AIDessa Voice AI Model:Trained Only on Speech
You learned to talk by listening. So did Dessa. Block's full-duplex speech model learned to hold real-time conversations from raw human speech alone — no text, no pretrained models. Listen to all 21 official demo recordings below, and test a voice AI right now with ElevenLabs v3 on Replicate.
What is Dessa?
Dessa is the most authentic full-duplex speech model ever built, according to Block, the company behind Square and Cash App. Talking with it feels like talking with a person — full of emotion, nuance, pauses, and a natural cadence.
Current voice AI "can just plain suck": it turns your voice into text and throws away an incredible amount of information text alone can't carry — how you feel, when you hesitate, when you laugh. Dessa was trained from scratch on nothing but human conversation: sounds. No text, no pretrained models. It learned to hold a real-time conversation simply by listening — as far as we know, the world's first AI you can have a realistic live conversation with that was trained only from human speech.
Because speech never repeats — say a sentence twice, exactly the same way, and the data still comes out different — training on raw audio is brutally hard. A single spoken word takes thousands of times more bytes than the same word in text. Yet the Dessa team figured it out, and trained the model on only one NVIDIA GB200 NVL72, learning language from hundreds of times fewer words than text-based LLMs.
The result: a model that laughs the moment something's funny — mid-sentence, the way people do — finishes your sentence when you blank on a movie title, gasps "oh my god" under its breath when surprised, and speaks up when you ask it to. Inference runs faster than real time, and the model runs locally on a MacBook Pro M4 Max or better.
Why build it? Three reasons: curiosity, better understanding between humans and AI, and safety. "AI keeps gaining power and independence, and it's still nearly deaf to the emotion in our voices," the team writes. "Imagine being in serious pain while a voice AI has no idea and just keeps going. That's the safety problem. Dessa is a step toward fixing it."
What makes Dessa different
Pure speech-to-speech
Audio in, audio out. No text pipeline in between — everything Dessa knows came from human speech.
Full-duplex conversation
Listens and talks at the same time. Knows when to cut in and when to backchannel ("yeah!", "got it!", "ooh").
Emotion, sarcasm & laughter
Natively hears and speaks emotion, sarcasm, laughter, a frustrated sigh, or a voice cracking with sadness.
Extreme data efficiency
Learned language from hundreds of times fewer words than text LLMs — trained on a single NVIDIA GB200 NVL72.
Runs locally
Inference runs faster than real time, locally on a MacBook Pro M4 Max or better. No cloud round-trip.
Safety-first research
Built partly to fix a real safety problem: AI that cannot hear emotion cannot know when a person is in pain.
Let me listen… 🎧
Every clip below is a live recording of a person talking to Dessa. Nobody steers it — Dessa picks what to say, when to say it and how, down to its own voice and personality. Play both speaker tracks in sync and read along with the transcript.
Talking about the release
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.
Remembering a movie
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.
An anniversary surprise
A surprise. Dessa gasps “oh my god” under its breath, then answers with what sounds like teary gratitude.
Expressive conversation
Huge range in intonation and speed.
Life advice
Here it sounds genuinely empathetic.
Dessa in conversation with GPT-Live-1
And here it talks to ChatGPT Live.
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.
More Dessa voice samples
Artist Dessa
Bathroom disaster
Hobbies talk
Lazy AI gaming
Marvel gaming
Noisy GPT live
In this demo, Dessa is talking with GPT-Live-1, a third-party AI model. Both voices are AI-generated.
Oh my god
Poor Elina
Refund
Ronaldo
Safety lab
Singing nonsense Spanish
Spontaneous AI
Wise Dessa
Zelda discussions
Audio samples collected from the official block.xyz/dessa page. More recordings will be shared by the team on X.
Try a voice model yourself
Dessa is waitlist-only for now — but you can test expressive AI voices today with the ElevenLabs v3 text-to-speech model, running on Replicate. Use the prompts below: they are real lines Dessa spoke in the demos above.
Try a voice model right now
Dessa itself is a private research preview (join the official waitlist). In the meantime, test the most expressive TTS model available — ElevenLabs v3, running on Replicate.
Voice controls
Higher = more consistent, lower = more expressive
How closely the output matches the chosen voice
How much to lean into the voice style
Speaking rate
Powered by the ElevenLabs v3 model on Replicate · ~$0.10 per 1,000 characters · 26 voices · 70+ languages · Output: MP3
The team behind Dessa
A (very) small AI research lab at Block, founded by former members of Dessa, the Toronto deep learning startup. The team has built deep learning models since 2012 and generative speech models for over eight years, and reports directly to Jack Dorsey. Led by Ragavan Thurairatnam.
Researchers: Louay Hazami, Hashiam Kadhim, Qi (Alex) Liu, Rayhane Mama, Aydin Polat, Marc Tyndel · Interns: Cara Ji, Jibran Iqbal Shah, Dvir Zagury-Grynbaum · Former interns: Aryan Sharma, Olivia Yan
Calling all safety researchers
A model that learned to talk from sound alone is new territory. Block is giving a small group of AI safety researchers early access to find the risks they haven't anticipated.
Apply for early access on the official site ↗Dessa voice model FAQ
What is Dessa, the AI voice model from Block?+
Dessa is a full-duplex speech-to-speech conversation model built by a small AI research lab at Block (Square). It was trained from scratch on human conversation audio only — no text, no pretrained models — and it learned to hold real-time conversations simply by listening. Block describes it as the first AI you can have a realistic live conversation with that was trained only from human speech.
How is Dessa different from regular text-to-speech and voice assistants?+
Most voice AI converts speech to text, processes it, and converts text back to speech. That pipeline strips away emotion, sarcasm, laughter, sighs, and natural timing. Dessa is a true speech-to-speech model: audio in, audio out. It natively hears and speaks emotion, laughs mid-sentence, backchannels ("yeah!", "got it!"), cuts in at the right moment, and talks and listens at the same time (full-duplex).
How was Dessa trained, and what hardware does it need?+
Dessa was trained on only one NVIDIA GB200 NVL72 and learned language from hundreds of times fewer words than text-based LLMs. Inference runs faster than real time and works locally on a MacBook Pro M4 Max or better.
Is Dessa open source? Can I use the Dessa API?+
Not yet. Dessa is a research preview. Block is giving a small group of AI safety researchers early access, and everyone else can join the waitlist at block.xyz/dessa. The team also said more demo recordings will be shared on X.
Where can I listen to real Dessa voice demos?+
This page hosts all 21 official demo recordings from block.xyz/dessa with synchronized dual-track audio (human + Dessa) and full transcripts. The original clips are also available on the official Dessa site.
How can I test voice AI models like Dessa today?+
While Dessa is waitlist-only, you can test the ElevenLabs v3 text-to-speech model (hosted on Replicate) in the playground on this page. It offers 26 voices, 70+ languages, and controls for stability, similarity, style, and speed — and it powers expressive voice output that previews where voice AI is heading.