Block released Dessa, a full-duplex speech model trained exclusively on human conversation audio. The research preview arrived on October 23, 2026. Unlike standard voice agents that rely on text-first stacks, Dessa processes audio input and produces audio output without using text or pretrained models.
TL;DR
- Runs on: A MacBook Pro M4 Max or better, locally, faster than real time.
- Performance: Demo-only so far. Live recordings show backchannels, mid-sentence laughs and natural interruptions.
- Best: Trained on 1 NVIDIA GB200 NVL72 rack (72 GPUs). Moshi used 1,016 H100 GPUs.
- Worst: 0 published benchmark scores. No weights, paper, API or price.
What is Block’s Dessa?
Dessa is a true speech-to-speech model. Audio in, audio out. According to Block’s project page, everything the system knows came from human speech.
It is full-duplex, so it listens while it talks. Block says it learned from real conversation when to cut in. It also learned when to backchannel with short cues like “yeah” or “got it”.
Block’s demo clips show Dessa laughing mid-sentence and getting louder when asked to speak up. One clip shows it talking with OpenAI’s GPT-Live-1. Block says nobody steered any clip.
How does speech-only training work?
Most voice systems chain three parts: speech recognition, a text model, then speech synthesis. The Block team argues that chain throws away emotion, sarcasm, laughter and sighs. Dessa hears and produces those signals directly.
Block has not published the architecture. Here is what it has disclosed:
- Training data: human conversation audio only. No text and no pretrained models.
- Compute: 1 NVIDIA GB200 NVL72, which links 72 Blackwell GPUs and 36 Grace CPUs.
- Data efficiency: Block says Dessa learned language from hundreds of times fewer words than LLMs need.
- Inference: faster than real time, on a MacBook Pro M4 Max or better.
The lab also explains why this is hard. Speech never repeats exactly. A spoken word takes thousands of times more bytes than the same word in text.
Kyutai’s Moshi, the best-known open full-duplex model, takes another route. Moshi starts from a text language model backbone and predicts time-aligned text tokens before audio tokens. Its model card lists 7 million hours of pretraining audio and 1,016 H100 GPUs.
How does Dessa compare with Moshi, PersonaPlex and GPT-Live-1?
| Feature | Dessa (Block) | Moshi (Kyutai) | PersonaPlex-7B-v1 (NVIDIA) | GPT-Live-1 (OpenAI) |
|---|---|---|---|---|
| Starting point | Speech only, no text, no pretrained model | Helium text LM backbone plus audio | Fine-tuned from Moshi (Moshiko) weights | Not disclosed; delegates reasoning to a backend agent |
| Parameters | Not disclosed | 8B (Hugging Face listing) | 7B (model card) | Not disclosed |
| Training compute | 1 GB200 NVL72 (72 GPUs) | 1,016 H100 GPUs | Not disclosed | Not disclosed |
| Training audio | Not disclosed | 7M h unsupervised, 2,000 h Fisher, 20,000 h synthetic | Fisher English, under 10,000 h | Not disclosed |
| Latency | Not disclosed (“faster than real time”) | 200 ms in practice, 160 ms theoretical | 0.170 s turn-taking, 0.240 s interruption (FullDuplexBench) | Not disclosed on model page |
| Weights and license | Not released | Open, CC-BY 4.0 | Open, NVIDIA Open Model License | Closed |
| Runs on | MacBook Pro M4 Max or better | GPU; MLX, Candle and PyTorch checkpoints | NVIDIA A100 / H100, Linux | OpenAI cloud |
| Access and price | Research preview, waitlist | Free download | Free download (gated) | API, $0.05 per minute voice layer |
What it means
For people making things, the lack of a public API means Dessa is not ready for immediate integration into production workflows. The system runs locally on high-end hardware, which is useful for developers testing voice interactions without sending data to the cloud. However, the absence of benchmarks, weights, or pricing details makes it impossible to evaluate performance against commercial alternatives or estimate costs. The waitlist access suggests Block intends to move from research to a controlled release, rather than offering open weights immediately.
