Block 发布纯语音训练的全双工语音模型 Dessa
Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech - MarkTechPost
Block 的 AI 研究实验室发布全双工语音模型 Dessa，以研究预览形式开放候补名单，只用人声对话音频训练，未使用文本和预训练模型。该模型可在 MacBook Pro M4 Max 或更高配置上本地运行，速度快于实时；训练使用 1 台 NVIDIA GB200 NVL72（72 块 GPU），目前未公布基准分数、参数量、权重、论文或 API。
正文
Add as a preferredsource on Google
Block’s AI research lab has released Dessa, a full-duplex speech model trained only on human conversation. Block announced it on October 23, 2026 as a research preview. Dessa is a speech-to-speech model from Block that listens and talks at the same time. Block says it used no text and no pretrained models. That is a different design from the text-first stacks behind most voice agents today.
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 it knows came from human speech.
TEXTTABLELAYOUTIs your PDF parser actually getting it right?omni_parse_bench · scores text, tables & layout · open sourceStar on GitHub→Sponsored
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 3 parts: speech recognition, a text model, then speech synthesis. 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?
Starting pointSpeech only, no text, no pretrained modelHelium text LM backbone plus audioFine-tuned from Moshi (Moshiko) weightsNot disclosed; delegates reasoning to a backend agent
ParametersNot disclosed8B (Hugging Face listing)7B (model card)Not disclosed
Training compute1 GB200 NVL72 (72 GPUs)1,016 H100 GPUsNot disclosedNot disclosed
Training audioNot disclosed7M h unsupervised, 2,000 h Fisher, 20,000 h syntheticFisher English, under 10,000 hNot disclosed
LatencyNot disclosed (“faster than real time”)200 ms in practice, 160 ms theoretical0.170 s turn-taking, 0.240 s interruption (FullDuplexBench)Not disclosed on model page
Weights and licenseNot releasedOpen, CC-BY 4.0Open, NVIDIA Open Model LicenseClosed
Runs onMacBook Pro M4 Max or betterGPU; MLX, Candle and PyTorch checkpointsNVIDIA A100 / H100, LinuxOpenAI cloud
Access and priceResearch preview, waitlistFree downloadFree download (gated)API, $0.05 per minute voice layer
Key Takeaways
Dessa is full-duplex and speech-to-speech, trained only on human conversation audio.
Block trained it on 1 GB200 NVL72 rack; Moshi used 1,016 H100s.
It runs locally on a MacBook Pro M4 Max, faster than real time.
Block published 0 benchmarks, no parameter count and no weights.
Access is a waitlist.
Michal Sutter+ postsBio ⮌
Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.
Michal SutterMirendil Launches Fern 0.1, Its First Model, With 10M-Token Context
$uv pip install omni-parse-bench✓text✓tables✓layout→ scored per pageStar on GitHub→Sponsored
RELATED ARTICLESMORE FROM AUTHOR
Baseten Releases joey-1, a 2B Decision Model at $0.01 per 1M Tokens
Prime Intellect Releases Prime Mega MoE: Trainable Fused MoE Kernel, 3.11x Faster Forward Pass
NVIDIA Releases Nemotron 3.5 Transcribe: An Open 30B MoE Model for 40-Language Transcription with ~4B Active Parameters