发布会·中文精读

将思想转化为语音 | Neuralink

Translating Thought To Speech | Neuralink

中文精读由 AI 基于原始视频与文稿改写,非官方翻译,可能存在偏差或错误;表述以原视频为准。说明

2026 年 1 月,一个叫 Ken 的男人做了一场开颅手术。医生在他头骨上开了个小孔,把一根比头发丝还细的电极线送进大脑。第二天,他出院回家了。

几周之后,他坐在镜头前,说出了这样一句话:

我在用思想和你说话。(Ken)

这句话没有经过声带,没有经过舌头和嘴唇,也没有经过任何一块该动的肌肉。他只是想了一下,机器替他把声音发了出来。

Ken 是一名 ALS 患者。这种病会一点一点切断大脑和肌肉之间的连接,最先被夺走的往往是声音。Neuralink 做的事,是绕开那条被切断的路,直接从大脑里把话读出来。更特别的是,替他说话的那个声音,是他自己 2020 年的声音。

一根被剪断的网线

先解释 ALS 到底坏了什么。

ALS 全称肌萎缩侧索硬化,它损伤的是运动神经元。大脑里负责下达运动指令的神经元还在,把指令送到肌肉的那条线路断了。用 Neuralink 团队在片子里的说法,这就像一根被剪断的网线:信号源还在发数据,接收端还在等,中间的线没了。

结果就是肌肉慢慢失去控制。走路会摔,手抬不起来,说话越来越含糊,最后连吞咽和呼吸都会受影响。

Ken 的妻子记得那些早期信号。他们曾经一起开租车公司,Ken 负责把客人从机场送到会场。她说那段时间开始注意到一些不对劲的地方,Ken 自己也说不清,只是"我试着把身体撑起来,我摔倒了,就是不对劲"。

声音的变化来得更晚一些。疫情期间,她发现 Ken 的声音开始不一样了。2024 年 1 月,确诊。

她说自己当时感觉整个世界压了下来,非常害怕。她已经失去过一任丈夫,她猜 Ken 那时最担心的事,是再一次把她一个人留在这个世界上。

这一段对话在片中只有几十秒,但它把后面所有技术细节的分量都定住了。镜头里没有极客在玩硬件,只有一个正在失去表达能力的男人,和一个害怕再次独处的女人。

说话比你想的耗体力得多

Ken 描述自己的处境时,说了一句很容易被忽略的话。

很多时候别人根本听不懂我在说什么,这太让人沮丧了。问题不只是肌肉不工作,还有疲劳。你没体会过就不知道说话要消耗多少能量,直到你连这点能量都拿不出来。这真的会夺走你的独立性。(Ken)

健康人说话几乎不耗力气,所以很难理解"说话是一种消耗"。对 ALS 患者来说,每一个音节都要用尽全力去挤,说上十几分钟就撑不住了。

他妻子补上了另一半代价:

我们家变得非常安静。ALS 带来的那种沉默,是我一开始完全没意识到、后来却感受最深的东西。(Ken 的妻子)

一个家里如果没人说话,房子会变。她说这话时语气很平,像在讲一件家具的事。

语音区就在手部运动区下面两三英寸

Neuralink 的第一款产品叫 Telepathy(心灵感应)。它记录的是运动皮层里一小块区域的神经活动,这块区域负责执行手和手臂的动作。四肢瘫痪的人想动右手,那里的神经元就会放电,设备把放电模式翻译成光标移动或者点击。

语音产品的位置稍微偏下一点。

再往下走大概两三英寸,就是另一个脑区,负责执行说话这个动作。(Neuralink 团队)

这个"两三英寸"是整件事的物理基础。大脑皮层的运动区有一张身体地图,手、手臂、脸、舌头、喉咙的对应区域挨得很近。伸手的指令和发"啊"音的指令,在地理位置上几乎互为邻居。所以做手部解码的团队,往旁边挪一点点,就能碰到发声的回路。

设备本身的能力是这套方案的底气。团队的说法是:

Neuralink 的设备给了我们别的设备给不了的东西,那就是同时从数千个通道做单神经元级别的记录。(Neuralink 团队)

这句话值得拆开。单神经元级别意味着电极足够近,能听见单个细胞在说什么,而不只是听一片区域的嘈杂平均值。数千通道意味着采样量足够大,大到能把复杂的发声动作拆成足够细的信号。语音是一连串快速变化的肌肉指令,含糊的平均值不够用。

为什么"想说话"还能被读出来

这里有个很多人会卡住的地方:既然 ALS 让肌肉不能动,那"想说话"这个动作还存在吗?

存在。至少在相当长一段时间里,ALS 损伤的是通路,不是指令源。大脑照常发出"我要说这句话"的电信号,只是这个信号到不了嘴唇和舌头。BCI 做的事就在这条断掉的线上架一座桥。

有了脑机接口,你完全绕开这条通路,可以直接从大脑读到设备上,去做你想做的事。(Neuralink 团队)

Ken 是怎么找到这个项目的?他妻子说,是在网上看到的。看到 Neuralink 开放了面向 ALS 患者的患者登记(Patient Registry),他当天就报了名。

先把 2020 年的声音存下来

语音产品有一个手部产品没有的难题:声音是有身份的。

你能读懂屏幕上滚动的字,但那不是你。你听到机器合成的语音说出你想说的话,那也不完全是你。人对自己的声音有非常强的认同感,声音消失,等于一部分自我被抹掉了。

所以 Neuralink 没有给 Ken 配一个通用音色,用的是他本人。

片头那句话交代得很清楚:Ken 在 2026 年 1 月成为第二位接受语音植入的参与者,他想着,设备替他说,而且听起来会像他本人,像 2020 年、声音还没变的那个他。片子里直接把这个声音叫作 OG Ken(元老版 Ken)。

技术上这叫语音克隆。做法是先拿到这个人健康时期的大量语音样本,让模型学会他的音色、语速、口音和发音习惯,再把大脑解码出来的语言内容送进这个音色里合成。解码负责"说什么",克隆负责"听起来是谁"。这两件事必须同时成立,产品才真正有意义。

那句被找回来的声音

片子最动人的一段不在技术演示里,在 Ken 的妻子回忆第一次听到设备开口的时候。

他们启动设备,他用 OG Ken 的声音告诉我他爱我。我太激动了,因为我用这个声音在电话里听过那么多年。(Ken 的妻子)

她和 Ken 的故事开头很有画面感。两人在 eHarmony(一个婚恋网站)上认识,连续发消息发了 19 个小时。那天结束时 Ken 问,能不能给你打个电话。她听到他声音的那一刻,就确定了。

算是听声音就爱上了吧。(Ken 的妻子)

所以她听到的不只是一段合成音频,是二十年前电话那头的那个人。

手术第二天他就回家了

往脑子里插电极这件事,很多人的印象还停在科幻片里。Ken 的实际经历是这样的:手术做完,第二天送回家。他妻子说整个过程顺得不可思议,两天之后你根本看不出他发生过什么事。

手术后他们在达拉斯住了一段时间,做随访和调试。

团队这边的节奏也在加快。片子拍摄时,他们提到上周刚做完下一台语音植入手术。流程是这样的:

  • 第一天上门,教用户用充电器、用 App
  • 引导用户尝试说一些句子,把这些数据采下来
  • 用采到的数据训练模型,把神经意图映射到实际词汇

注意第一步和第二步之间的关系。训练数据是用户自己"说"出来的,只不过说的是口型,不是声音。设备记录下他想说某个词时大脑的样子,模型才知道以后该怎么翻译。

第一次训练:雨落在西班牙

调试那天的现场,团队让 Ken 用一半的速度做口型,一句一句来。

他们选了那句经典的英语发音练习句:The rain in Spain stays mainly on the plain(雨主要落在西班牙平原上)。这句话出自音乐剧《窈窕淑女》,用来练发音再合适不过,因为它把英语里最容易糊在一起的几个元音全塞进去了。

Ken 慢慢做口型,团队在旁边看数据。

然后几乎是从零直接跳到了一百,模型突然就通了。(Neuralink 团队)

屏幕上逐词跳出:The、rain、in、Spain、stays、mainly、on、the、plain。Ken 自己都愣了一下,说了一句:

这就是我说的。(Ken)

团队问他刚才有什么不一样。他说:

脑子里那个东西,我更用力了,意图更强。(Ken)

团队的回答解释了一切:

那样会更好地调制你的神经信号。(Neuralink 团队)

"更用力地想"是一门要练的手艺

这句对话里藏着脑机接口最容易被低估的部分:用户不是被动被读取,用户要学着配合。

同一个意图,想得敷衍和想得明确,产生的神经信号强度不一样。信号强,模型就好解码。所以每个使用者都要摸索一套自己的"发力方式":嘴型要不要动,动多少,注意力集中在哪个感觉上。

团队的说法很直白,用户戴上设备以后,需要花时间玩一玩自己的行为,看哪种方式效果最好。

片子还拍到 Ken 用这套系统干别的事。他对着设备说"Siri,把每个房间的灯打开",然后灯亮了。这一步看着简单,其实闭环已经完整了:大脑发出指令,设备解码成语音,语音助手识别,家电执行。他靠想,控制了自己家里的物理环境。

他们还一起玩牌。Ken 说"我要三张",说"我这轮结束了"。这些日常琐碎的对话,是产品能不能用的真正标准。

还有一段是他在念《彭特兰海盗》里那首出了名难念的快歌:"我是现代少将的典范"(I am the very model of a modern major general)。这首歌要求极高的语速和清晰度,用来测试极限。

目标不是慢慢说,是实时说

现在的形态还是一句一句来:Ken 意图说话,设备解码,语音输出。中间有延迟,需要他说完一段再等结果。

Neuralink 给出的下一步很明确:

我们会继续提升传感器的质量和数量。我们要建一个系统,直接从大脑到声音,实时完成。Ken 会得到即时反馈,并且能控制这个声音。(Neuralink 团队)

实时这两个字,是产品从"辅助沟通"走到"正常说话"的那条线。有延迟的对话只能轮流发言,实时对话才能插话、能打断、能开玩笑。人对交流的耐心很短,慢半拍的系统会被自然地排除在社交之外。

要做到实时,两头都得升级:硬件上要有更多更好的电极,软件上要压缩解码延迟。片子没给时间表,只说了方向。

另一个方向是个性化。片中提到每个用户会有自己的标识颜色,Ken 的是蓝色。这种细节看起来像产品经理的坚持,但它指向同一件事:这套系统属于使用者本人,不属于医院。

一个患者如何变成先行者

Ken 的妻子说,Neuralink 给他们的不只是 Ken 的声音,还有一种被推着往前走的力量,让 Ken 有机会成为一个先行者。

Ken 自己对这件事的表述更朴素。他说,如果不参与,那会是对他这个人的彻底浪费。他说自己最在意的部分是:

觉得自己在为别人做点好事,这是所有事情里最好的那件。(Ken)

他给后来者留了一句话:

你是自己这艘船的船长。用好生活给你的东西。(Ken)

片子结尾,他说了那句"我终于能和你说话了"。这句话的对象是他妻子。

"从减少人类苦难开始"

片子开场有一段话,交代了 Neuralink 的优先级:

我们从减少人类苦难开始。我们的第一批产品叫 Telepathy,它让失去身体控制能力的人能够和计算机沟通,也能实现概念层面的、双方同意的意念交流。(Neuralink)

"减少人类苦难"这句话在这类公司里经常被当成口号。放到 Ken 的故事里,它变得非常具体:一个说不出话的人重新能说话,一个安静下来的房子重新有声音,一个担心再次独处的女人不用再担心。

后半句提到的"概念层面的、双方同意的意念交流",是更远的东西。它的前提是信号带宽足够高,高到能传输词句之外的东西。那是很多年以后的事,但方向已经写在了第一版产品的命名里。

这件事对普通人意味着什么

现在能确认的事实有这么几条。

语音植入已经在 ALS 患者身上跑通了完整链路:手术、恢复、数据采集、模型训练、日常使用。手术的侵入性比想象中小,Ken 第二天就回家了。声音的身份问题有解,靠术前语音样本做克隆。瓶颈卡在"读得够不够快、够不够稳"这一头,"能不能读"早就有了答案。

Ken 身上那个反直觉的事实更该被记住:当一个人失去说话的能力,他想说话的意愿一点没减少。设备之所以有效,是因为大脑里那个"我要说这句话"的信号一直在,只是没人接。

技术在这里做的事情,说到底是把一条断掉的线重新接上,谈不上创造一种新能力。

如果你关注这个领域,下面几件事现在就可以做:

  • 家里有 ALS、渐冻症或严重运动功能损伤的患者,去 Neuralink 官网的患者登记(Patient Registry)看看是否符合入组条件,Ken 就是看到开放登记当天报的名,早排队不吃亏。
  • 如果确诊了会影响说话的疾病,尽早、尽可能多地录下自己正常状态下的语音,读文章、聊天、唱歌都行。语音克隆的原料是健康时期的声音,等到声音变了再录,就来不及了。
  • 关注这个领域的进展时,别只看通道数、电极数这类硬件指标,重点看延迟和实时性。从逐句输出到实时对话,是这类产品真正跨过门槛的那一步。
  • 和患者家属聊天时,别只问身体情况。问一句"家里还热闹吗",可能比什么都重要,因为沉默是这类疾病最容易被忽略的症状。
  • 对这项技术保持一点耐心,也保持一点怀疑。它现在能做的是一件很窄但很重的事:把声音还给一个正在失去声音的人。别急着把它推广成万能方案,也别因为它不够万能就否定它。
YouTube 平台字幕(自动或人工),可能有识别错误,人名与数字请以视频为准。已按语句合并成段落,并去掉 um、uh 等口头填充词。 文稿来源:YouTube(自动/人工字幕)。

0:03Hello. There we go. I'm talking to you with my mind. We're starting off with reducing human suffering. Our first products are called telepathy, and [music] that enables someone who has lost the ability to command their body to be able to communicate with a computer, and the ability to do conceptual consensual telepathy.

0:38In January of 2026, Ken became the second participant [music] to receive a Neuralink implant for voice. He will be able to think, and his device [music] will speak for him. And it will sound like him. It will be OG Ken [music] from way back in 2020 before his voice changed. I'm not certain about whether soulmates exist or not, but if they do, you're mine. I took one look at your dating profile, and I was like, "Wow." Hubba hubba. I was smitten. [laughter] Ken and I had a whirlwind romance.

1:20We met on eHarmony, and [music] we actually texted for 19 hours straight. At the end of that day, he says, [music] "You know, can I call you?" And when I heard Ken's voice, that was it. It was love at first sound, I guess, is kind of the way it is. After we met, we actually started a black car business. He drove people from the airport [music] to meetings.

1:44It was at that point where we had kind of started noticing things changing. I didn't notice that, but I tried to raise these up on the ground. I was falling over. It just wasn't normal. And during COVID is when we noticed his voice starting to change. In January of 2024, [music] we received the news that Ken had ALS.

2:10I remember feeling [music] kind of the whole weight of the world and it was really scary. I had already [music] lost a husband. I think Ken was worried that he was going to leave me alone again. That was hard for you. I like to God when the weather was going to do this. And then he would go shopping anymore.

2:35And then he can't get up. A lot of the time just people don't understand my speech. It's really frustrating. It's not only just the muscles not working, but it's the fatigue. You don't realize how much energy it [music] takes to talk until you don't have that energy to waste. And it really takes away [music]

3:01independence. Our house got really quiet. [music]

3:09It's the silence that comes with ALS was kind of one of the things I didn't realize when it started was going to be one of the things I noticed the most. The Neuralink device gives us something that no other device can give us, which is in single neuron recording from thousands of channels simultaneously.

3:28The Telepathy products is basically recording the neural activity from the small area in the motor cortex that involve in execution of hand and arm movements. But if we go only about 2 or 3 inches below, there's another brain area that's involved in execution of speech. We first heard about Neuralink online.

3:48And when he saw that they had a patient registry open for people with ALS. He signed up that day. When you have ALS, you have neuron degeneration. You're unable to transport the signal from your brain to your muscles. It's like a severed internet cable. [music]

4:07And with a brain-computer interface, you bypass this pathway completely, and you can read directly from your brain to a device to do the things you want to do. Like most ALS patients, Kenneth [music] was losing his ability to speak. It was slow. It was tiring and difficult, and it's something that is progressively getting worse.

4:28And so, the real value of the brain-computer interface is that it would give Kenneth [music] his voice back. He'd be able to speak comfortably for hours on end by no longer having to speak out loud, but rather just intend [music] to move his mouth silently in a way that's extremely comfortable.

4:47He was so excited, and I could tell he was excited. I could tell he was a bit nervous. Only cuz I know that was the surgery is up. Please wait. My capacity is lower. Otherwise, I would come home. Ken had surgery, and they sent him home the next day. It was such a smooth process. You would have never known, sort of 2 days down the road, that anything ever happened.

5:17We stayed in Dallas for a time after the surgery. We just did our second voice participant surgery last week. Today will be our first day with BCI. We'll go in, show him how to use his charger, how to use his app, and we will guide him for some sentences to attempt to say, and we'll use that data to try to map neural intent to the actual words.

5:44Yes, this is it. Awesome. They have that. So, when a user gets a brain computer in your face, it's sometimes important for them to play around with their behavior to see what works best. What we did was Kenneth slowed down a little bit and tried to enunciate kind of the individual sounds of his sentence.

6:07And really like out of nowhere, we went from zero to 100 with the model. The rain in Spain stays mainly on the plain. Ow. The rain in Spain stays mainly on the plain. That's exactly what I said. Hey.

6:42in. Hello. Nice to see you, Charlie.

6:44see you, guys. Good morning. Come on in. Good morning, Ken. Good to see you. We're still in an early phase here, but Kenneth is helping us build the future. We'll start with some data collection just to collect some sentences. After that, we'll train some models and and get going.

7:03We'll do these at half speed with mouthing [music]

7:06and just continue in that direction. Siri, turn on lights in every room. I think this is awesome. The goal is like we want him to be able to simply just intend to move his mouth and for our BCI to decode his speech. There There we go. I'm talking to you with my mind. Nice work. What was different there?

7:50And the thing in my head. Putting more effort in, more intention. It's modulating your neural signals more. I am the very model of a modern major general. Neuralink has given us not only Ken's voice back, but it's really given [music] us a boost. It's given us the chance for Ken to kind of be a pioneer.

8:22What excited me, they fired it up and he told me that he loved me in OG Ken voice. So emotional for me to hear it the way I heard it on the phone all those years ago. Yeah. Something else. I am finally able to talk to you. [music]

8:48We're going to continue advancing the quality of the sensors, the number of the sensors. We want to build a system that goes directly from the brain to voice in real time. And Kenneth will have instantaneous [music] feedback and control of this voice. Everyone gets his own [music] color.

9:08Will be blue. I'm going to pay three to play [music] search for life. I will take three. You going to do that? Yes, please. I'm done with my turn. We're better together. We'll make this thing better. Maybe I can't go ahead this. But again, participate in features. Again, give feedback.

9:46It will give us back a voice. And a voice is a very important thing [music] to have. It would be a total waste of who I am. What did you all the good you leave out? The world. The world turned out. Good to be with the world. And that's the number one reason I'm doing feeling I'm doing good for others.

10:25And that is the best thing of all. You are the captain of your own ship. Make the most of what you're given in life. [music]

资源与出处

相关视频

同频道、同主题或相邻年份