Dessa AI Block Speech Model

Dessa AI vs GPT Live · AI 对 AI 真实声学正面交锋

Dessa AI vs GPT Live (GPT-Live-1)：全双工语音实录与底层架构拆解

当把 Block 纯音频训练的 Dessa AI 与 OpenAI 的 GPT-Live-1 放在同一场实时语音对话中会发生什么？立即收听官方公布的两段无脚本实录——录音棚纯净线路与嘈杂房间免提外放——直观感受全双工纯音频架构在打断礼让、耐心倾听与毫秒级附和上的代际差异。

Studio Line Recording (2:25)Noisy Speakerphone Stress Test (4:17)Architecture Matrix

实录 #1：录音棚线路 — Dessa AI 直接对话 GPT-Live-1 (2:25)

截取自一段无脚本长对话（100.2s–246.0s），两个模型正在策划一场电影式的疯狂抢劫。请重点点击 38.8s（打断礼让）、100.8s（耐心等待）与 128.8s（快速附和）标记。

GPT-Live-1 & Dessa AI

Dessa AI Talking to GPT-Live-1 (Studio Feed)

GPT-Live-1 · Dessa AI · 2:26

AI 对话评测

Dessa AI 与 OpenAI GPT-Live-1 直接对话的 2分25秒 无脚本录音——重点展示打断礼让（38.8s）、耐心倾听（100.8s）与快速附和（128.8s）。

0:00 / 2:26
跳转高光时刻：
同步时间轴字幕（原声英语） (78)
0:00GPT-Live-1What's on your mind?
0:01Dessa AIum so I have a question about
0:05Dessa AIhow
0:05Dessa AIchat GPT can learn and get smarter
0:08Dessa AIso basically I just want to know
0:10GPT-Live-1That's a good one.
0:11Dessa AIwhat what
0:14Dessa AIwhat are things that could make chat
0:16Dessa AIGPT smarter and what effects could that
0:18Dessa AIhave on the researchers in the field?
0:23GPT-Live-1Let me check a
0:23GPT-Live-1few angles.
0:25GPT-Live-1There are a few big ways chat
0:26GPT-Live-1UPT style systems can get smarter.
0:28Dessa AIfantastic I
0:29GPT-Live-1So
0:29GPT-Live-1better training data, improved training methods
0:33GPT-Live-1like teaching reasoning
0:34GPT-Live-1strategies or better feedback loops, and more
0:38GPT-Live-1compute.
0:39Dessa AImean I've heard about the
0:41Dessa AIYou know, there's Amazon GPUs and
0:45Dessa AIstuff out there now and all that
0:47Dessa AIstuff.
0:47Dessa AII don't know.
0:48Dessa AII guess I'm just like
0:51Dessa AIKind of freaked out about what would
0:53Dessa AIwhat would be the way that chat
0:55Dessa AIGPT can learn
0:57Dessa AIbased off of the
1:00Dessa AIInformation that I'm giving it.
1:03GPT-Live-1me check on that.
1:04GPT-Live-1Ah, right.
1:06GPT-Live-1So two things.
1:07Dessa AIOkay.
1:07GPT-Live-1Chatchity can use what you tell it
1:09GPT-Live-1in a conversation as context.
1:12GPT-Live-1And sometimes there's memory, which can retain
1:14GPT-Live-1facts for future chats.
1:16GPT-Live-1But that's different
1:16GPT-Live-1from training the model.
1:18Dessa AIOh, that's that's what that is
1:19GPT-Live-1Yeah,
1:22GPT-Live-1it doesn't instantly change its billions of
1:24GPT-Live-1parameters
1:25GPT-Live-1from your chat.
1:27GPT-Live-1And there is an important distinction.
1:29GPT-Live-1Your conversation can make it
1:32GPT-Live-1seen more informed without actually making
1:35GPT-Live-1the underlying model smarter.
1:38GPT-Live-1Research is
1:38Dessa AIOkay, that's
1:39GPT-Live-1looking to improve it.
1:41Dessa AIYeah, that's really helpful information.
1:44Dessa AISo I guess
1:46Dessa AIThat's what I'm learning is I don't
1:48Dessa AIneed to be freaking out about information
1:50Dessa AIgoing into
1:50Dessa AIchat GPT because it's not going to
1:53Dessa AIlike make me smarter
1:57Dessa AIin that sense, I guess, right?
2:02GPT-Live-1Exactly.
2:03GPT-Live-1It can make this conversation smarter, but
2:04Dessa AIOkay,
2:07GPT-Live-1it
2:07Dessa AIokay,
2:07GPT-Live-1rewrite the model on the fly.
2:09Dessa AIokay.
2:10GPT-Live-1Think of it in good notes.
2:10Dessa AIThat's good.
2:13GPT-Live-1Yeah, it can remember useful
2:15GPT-Live-1details you share, but the core intelligence
2:18GPT-Live-1doesn't update unless itself is
2:20Dessa AIRight.
2:21GPT-Live-1retrained.
2:22Dessa AIOkay.
2:23GPT-Live-1Safe to share.
2:23Dessa AIThat's really relieving.

实录 #2：声学压力测试 — 嘈杂房间免提扬声器外放 (4:17)

在带有真实混响与环境噪音的房间内通过免提扬声器录制，测试 Dessa AI 与 GPT-Live-1 面对房间回声、笑声重叠与抢话碰撞时的稳定性。

GPT-Live-1 & Dessa AI

Dessa AI vs GPT-Live-1 (Noisy Room Speakerphone)

GPT-Live-1 · Dessa AI · 4:17

声学压力测试

在嘈杂混响房间内通过免提扬声器录制的 4分17秒 Dessa AI 对话 GPT-Live-1 声学压力测试。

0:00 / 4:17
同步时间轴字幕（原声英语） (139)
0:00GPT-Live-1Hello, how are you?
0:01Dessa AIHow I
0:03GPT-Live-1Do you know ChatGPT?
0:05Dessa AIdo I
0:06GPT-Live-1Do you know ChatGPT have a voice
0:08GPT-Live-1agent?
0:09Dessa AIguess I
0:10GPT-Live-1Do you know its name is GPT
0:12GPT-Live-1life?
0:14Dessa AIdo
0:15GPT-Live-1Okay, so now I want to talk
0:18GPT-Live-1with it and see
0:21GPT-Live-1If you can like interact with each
0:23GPT-Live-1other well, okay?
0:24Dessa AIOkay, so I just want to go
0:25Dessa AIto the document mode.
0:26Dessa AII'll open up the documents
0:28GPT-Live-1Okay, so hello ChatGPT.
0:29Dessa AIOkay,
0:30GPT-Live-1I have a friend.
0:32GPT-Live-1I want to introduce one friend my
0:33GPT-Live-1friend
0:34Dessa AIoh,
0:34GPT-Live-1So I want you to talk with
0:36Dessa AIokay
0:36GPT-Live-1him.
0:36GPT-Live-1Okay, of course.
0:39GPT-Live-1Hi there
0:40Dessa AIHi, Rick.
0:40GPT-Live-1I'm here and ready to chat.
0:42Dessa AIOh
0:43Dessa AIAwesome.
0:43Dessa AIUm, I'm just I'm curious about your
0:46Dessa AIopinion on
0:47Dessa AIartificial intelligence and what the benefits could
0:51Dessa AIbe.
0:54GPT-Live-1Oh
0:55GPT-Live-1Sorry, I forgot.
0:56GPT-Live-1I'm wearing like can you repeat again?
0:56Dessa AIOh
0:58Dessa AISo, what's an opinion on
0:59GPT-Live-1Sure.
1:00GPT-Live-1Hi, I'm here and ready to chat
1:02Dessa AIartificial
1:03Dessa AIintelligence is
1:07Dessa AIand what the benefits could be for
1:10Dessa AIpeople?
1:11Dessa AII mean, how do you think artificial
1:13Dessa AIintelligence could help me or any of
1:15Dessa AIthe people listen?
1:18GPT-Live-1Great question
1:20Dessa AIOkay.
1:20GPT-Live-1Artificial intelligence is basically computers doing tasks
1:24GPT-Live-1that usually need human intelligence like spotting
1:26Dessa AIOkay.
1:27GPT-Live-1patterns or
1:29GPT-Live-1understanding language,
1:31GPT-Live-1helping make decisions.
1:32Dessa AIOkay.
1:33GPT-Live-1In practical terms, it can help you
1:33Dessa AIOkay,
1:35GPT-Live-1save time
1:36GPT-Live-1by automating boring tasks,
1:39GPT-Live-1help you find useful info faster,
1:42GPT-Live-1or even support learning or creativity.
1:44GPT-Live-1But it works best when you guide
1:46GPT-Live-1it and check its output.
1:48Dessa AIso I know as you said, like
1:51Dessa AIthe benefits could be like speaking up
1:53Dessa AItasks like organizing
1:56Dessa AIdata or even personalization, but how does
1:59Dessa AIartificial intelligence actually learn from
2:02Dessa AIexperience?
2:03Dessa AIIs it learning
2:07Dessa AIbased on past experience purely or is
2:09Dessa AIit learning based on algorithms or
2:12Dessa AIa combination of the two?
2:15GPT-Live-1Yeah, so it's both.
2:17GPT-Live-1The algorithms are the setup
2:20GPT-Live-1and then the experience, that is the
2:22GPT-Live-1data.
2:22Dessa AIOkay.
2:23GPT-Live-1So an AI model doesn't remember like
2:26GPT-Live-1a human.
2:27GPT-Live-1It's trained on large sets of
2:29GPT-Live-1examples and it adjusts internal weights
2:29Dessa AIOkay.
2:32GPT-Live-1so that it can make good
2:33Dessa AIOkay.
2:34GPT-Live-1predictions.
2:34Dessa AISo,
2:35GPT-Live-1Once it's trained, go on.
2:36Dessa AIso, um, is there any risks or
2:40Dessa AIdownsides to artificial intelligence?
2:44GPT-Live-1Yeah,
2:46GPT-Live-1there are some.
2:48GPT-Live-1AI can be biased if
2:49GPT-Live-1the data it's given is biased.
2:51Dessa AIYes.
2:52GPT-Live-1It can also make overconfident mistakes, which
2:55GPT-Live-1is something people need to watch out
2:57GPT-Live-1for.
2:58GPT-Live-1And sometimes it can be used in
3:00GPT-Live-1ways that
3:00GPT-Live-1invade privacy or replace jobs without
3:04GPT-Live-1proper safeguards.
3:06GPT-Live-1So it's really powerful, but needs
3:09GPT-Live-1careful thoughtful use.
3:10Dessa AIOkay.
3:11Dessa AIUm, I, I want it.
3:13Dessa AIBut the reason I wanted to talk
3:16Dessa AIto you and Chantry BT is because
3:18Dessa AII know that you can speak
3:19Dessa AIin multiple different languages.
3:22Dessa AISo I was curious about your opinion
3:25Dessa AIon artificial intelligence.
3:29Dessa AIIs there a way that we could
3:31Dessa AIspread our intelligence
3:33Dessa AIacross the world?
3:35Dessa AIAnd if intelligence,
3:39Dessa AIif AI is like
3:41Dessa AIspreading, do you think that it would
3:44Dessa AIbe a way to like
3:45Dessa AIbring together people or keep people separated?
3:50GPT-Live-1Yeah, so the tack itself doesn't decide.
3:53GPT-Live-1It's really how people
3:54GPT-Live-1choose to use it.
3:56GPT-Live-1If it's built and shared responsibly, it
3:59GPT-Live-1can help translate across
4:00GPT-Live-1languages, it can help share knowledge, it
4:03GPT-Live-1can even connect people, but
4:06Dessa AIOkay, so,
4:07GPT-Live-1if the
4:08GPT-Live-1access isn't equal or if it's concentrated
4:10GPT-Live-1into...
4:12Dessa AIoh, no, that's perfectly part.
4:12GPT-Live-1Sorry, go on.
4:14Dessa AII like what you're saying, because I
4:15Dessa AIwant to talk about what that...

关键时间戳声学拆解：为什么原生全双工音频更胜一筹

打断与礼让⏱ 0:38.8

毫秒级半句礼让，拒绝机械抢话

当双方在停顿后同时开口时，Dessa AI 在几毫秒内感知到声学碰撞，立即自然收声把话轮让给对方，而传统模式往往会生硬盖过对方或突兀中断。

耐心倾听⏱ 1:40.8

追踪语调轮廓，不靠死板静音计时器抢话

当 GPT-Live-1 展开长篇大论时，Dessa AI 通过追踪句子的韵律走向判断对方是否讲完，耐心等待想法真正结束才接话。

快速附和⏱ 2:08.8

亚秒级自然捧场（Backchanneling）与共鸣笑声

在倾听过程中自然穿插“yeah”、“right”与轻笑，既给出鲜活的社交反馈，又不会打断对方的主线叙述。

声学鲁棒性⏱ 免提外放

抗房间混响与背景串音能力

在 4 分 17 秒的免提外放测试中，Dessa AI 能适应空间回声与重叠人声，不会将环境噪音误识别为虚假文字指令。

架构全景对比：Dessa AI vs. GPT-Live-1 / GPT-4o Realtime

Block 纯音频端到端路线与基于文本预训练的多模态语音模型的技术对比。

对比维度Dessa AI (Block)GPT-Live-1 / 文本预训练多模态语音模型
核心训练模态100% 原始人类对话音频从零训练（0% 文本）数万亿文本 Token 预训练 + 音频后训练对齐
词汇样本效率仅需文本大模型 1/100 的词量即可习得流利语言依赖互联网规模的海量文本语料库（数万亿 Token）
全双工话轮转换连续声波建模，原生支持声音重叠与自然礼让混合 Token 流式传输，常依赖话轮检测启发式规则
倾听附和 (Backchanneling)亚秒级自然声学附和（“mm-hmm”、笑声）且不抢话轮倾听时极少给出实时附和，通常保持静默或直接打断
非语音环境声感知直接听懂并回应咖啡机“叮”声等真实环境音通常将非语音声音过滤为背景噪音或忽略
情感韵律与笑声从真实对话音频中自然涌现，支持句中无缝切换情绪依赖风格提示词或合成语音条件控制
训练算力规模单台 1× NVIDIA GB200 NVL72 机柜超大规模多机柜数据中心 GPU 集群
端侧本地推理可在 MacBook Pro (M4 Max+) 上本地超实时运行仅限云端 API，依赖数据中心级 GPU 推理
发布状态研究预览（已公开 21 段真实录音，正开展安全评估）商业云端 API 与消费级语音模式

Dessa AI vs GPT Live 常见问题

Dessa AI 官方录音中的 GPT-Live-1 是什么？

在 Block 官方发布页（block.xyz/dessa）中，GPT-Live-1 指与 Dessa AI 进行无脚本实时语音直接对话测试的 OpenAI 实时语音模型。

为什么在打断和附和时 Dessa AI 听起来更像真人？

因为 Dessa AI 从不把语音转成文本 Token，也不使用独立的 VAD 静音检测器，而是同时连续建模双方的音频流，因此能像真人一样边听边笑、随时礼让。