Block releases Dessa, a full-duplex model trained purely on speech, which can run locally in real time on a MacBook Pro M4 Max
English text is being prepared.
Block, founded by Twitter co-founder Jack, has released Dessa, a speech-to-speech model trained from scratch purely on human speech, with no text and no pretrained base, supporting full duplex so it can listen and speak at the same time. The model claims inference faster than real time and can run locally on a MacBook Pro M4 Max and above, and the entire model is said to have been trained on a single NVIDIA GB200 NVL72. The team says its goal is to let the model understand tone and emotion, avoiding the safety risk of continuing to respond without perceiving the other person's distress.
Why read this
AI score 76
The original text sets out three points—training purely on speech, full duplex, and local real-time operation—from which one can infer the approach of a speech model that bypasses the text intermediate layer.
Topics
View original postWe believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. https://t.co/r0qFLdvd6c
View original
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. https://t.co/r0qFLdvd6c
Body
Twitter 联创 Jack 创办的 Block 发布 Dessa：纯语音训练的全双工会话模型，听得见情绪，Mac 本地实时跑
Dessa 是 Block 小团队从零用纯人类语音训练的 speech-to-speech 模型：无文本、无预训练底座，全双工边听边说，声称推理快于实时，可在 MacBook Pro M4 Max 及以上本地跑；动机是自然交互、理解情绪，以及「听不见痛苦还继续聊」的安全风险。 https://t.co/sOZgeLlnTa
# 它是什么
定位：目前据称首个仅从人类语音学会真实实时对话的 AI；交互目标是带情绪、停顿、自然节奏，像跟人说话。
技术形态：音频进、音频出；全双工（同时听与说）；会插话、会 backchannel（yeah / got it / ooh）；音色与个性由模型自选，演示为真人与 Dessa 的现场录音（字幕后加），无人实时操控。
演示侧重：句中即笑、被要求大声就变大、接上对方想不起的电影名、惊喜时压低嗓音「oh my god」、共情式建议、甚至与第三方 GPT-Live-1 互聊（双方均为 AI 声，GPT-Live-1 开发方与 Block 无关）。
# 为何不做「语音→文本→LLM」
现行语音 AI 常把声音压成文字，丢掉语气、讽刺、叹息、哽咽等文本带不走的信息；最强模型也「听不见你怎么感觉」。
团队认为这既损体验，也有安全隐患：人在剧痛时系统毫无察觉仍继续应答。
# 训练与算力叙事
从零、只靠对话声音学习；相对 LLM，用的「词量」少数百倍（文中也写：LLM 还要学 C++、Angular 等）。
难点：口音、停顿、呼吸、讽刺、重音改变语义（同句重读不同词含义全变）；语音几乎永不重复；一词字节量远高于文本。
声称整模型在 一台 NVIDIA GB200 NVL72 上训成。
# 为什么做（Wait but why）
好奇：婴儿不识字也能学说话，模型能否同样。
更好的人机理解：语音是最自然界面，现有方案浪费它。
安全：能力与自主性上升时，对情绪几乎「耳聋」是风险；Dessa 是朝「听得见情绪」迈的一步。