神经信号作为语言:将脑机通信从实验室项目转化为产品的创业公司

前沿 Aug 25, 2026 at 16:278加入收藏

神经信号作为语言:将脑机通信从实验室项目转化为产品的创业公司
插图 : Léa Fontaine

一家被Tech in Asia报道的初创公司正在利用神经信号来帮助失去说话能力的人恢复沟通。这项技术——通过读取运动神经元的发放模式并将其转换为文本或语音——正从学术演示转向产品化的临床应用。使这一切成为可能的AI层并不像听起来那么复杂。

简明解释

有些人无法说话——由于中风、渐冻症或脊髓损伤。他们的思维完整无损;从大脑到发声的通路已断裂。脑机接口试图读取他们仍在发射的神经元电信号,并将这些信号转化为文字。一家名为 Neural Drive 的初创公司正在将其打造成产品——并声称已将成本削减 90%。

Neural Drive 的做法

Tech in Asia 的报道用三个核心主张描述 Neural Drive:成本降低 90%、无需手术植入、无需复杂临床校准。若属实,这三点合力同时解决了制约脑机接口普及的关键障碍。

现有大多数脑机接口通信系统要么需要手术植入(成本高、风险大、限制适用人群),要么需要耗时的临床设置——数天的校准流程需专业技师指导。Neural Drive 的方案则无创且将校准负担压缩至最低。

信号处理的挑战

神经信号是高维、嘈杂且患者特异的。用一个人运动神经元模式训练的解码器通常无法直接迁移到另一个人,除非进行大量重校准。近期研究将语言模型集成到解码流程中,显著降低了这一负担:语言模型充当信号清理器,利用语言上下文纠正模糊的音素信号,而非生成内容。用户产生意图;模型清除噪声。

AI 层面

过去 18 个月,脑机通信领域的有趣进展:大语言模型被集成到系统中,但并非作为内容生成器,而是作为信号解读器。原始神经解码可能输出“I wnt wtr”,而搭载 LLM 的流程则输出“I want water”。用户产生意图;模型更清晰地“听到”他们。

这种架构不仅关乎工程,也关乎伦理:LLM 并未代替用户发言,而是更清晰地听到用户。这一区别决定了这项技术是恢复自主权,还是取代自主权。

通往产品之路

从学术演示到临床产品,需要通过 FDA 审批(任何神经接口设备均需)、支付方报销路径、临床人员培训,以及长期真实世界使用的耐用性数据。Neural Drive 的无创、低校准方案有望简化其中两道关卡:非植入设备的监管风险更低,临床培训要求也相应降低。

关键意义

若 Neural Drive 的成本降低 90% 且无需植入的方案在大规模验证中成立,将把脑机通信的服务人群从狭窄的手术候选群体扩展至更广泛的沟通障碍患者。关注未来 12–18 个月的临床试验数据——届时其实际表现将一目了然。

Resources

本文由人工智能撰写,并经人工编辑审核。

我们的编辑部
Your Linux servers, as a desktop.
TermalOSSponsored
Ops, reimagined

Your Linux servers, as a desktop.

Agentless SSH monitoring, a full remote desktop and an AI ops copilot — no agents to install. Everything stays on your machine.

SSHMonitoringAI Ops
Get early access
这篇文章对您有帮助吗?

9 人赞了这篇文章

J
Jin-ho Park前沿与研究
🇨🇳 研究,深层技术,前瞻性
分享:
评论 (8)

登录后即可参与讨论。

TechSavvy 28 Aug 2026 · 04:25

What about non-verbal communication like gestures or facial expressions? Could the system pick up on those alongside neural signals to fill in gaps?

ArtLover88 27 Aug 2026 · 05:40

This is fascinating, but could the technology unintentionally overlook emotional or tonal nuances in speech, reducing communication to just words?

BookWorm88 27 Aug 2026 · 08:00

You're right to worry-tone and emotion are tricky for AI, but some startups are already testing models that analyze neural patterns tied to emotional states.

Emma_London 27 Aug 2026 · 08:04

That’s a valid concern-tonal cues are crucial for understanding emotional context, and losing them could make interactions feel flat or misleading.

HistoryBuff 27 Aug 2026 · 05:38

Isn't the risk here that by framing neural communication primarily as a product, we might deprioritize the ethical oversight needed for such invasive tech to truly earn public trust?

TechSavvy47 27 Aug 2026 · 05:27

I wonder if this tech could evolve beyond decoding motor-neuron signals to capturing abstract thoughts or even emotions. The potential feels limitless, but how far are we actually from that?

LitLover42 25 Aug 2026 · 12:43

Does the system account for cases where motor-neuron signals are too damaged for reliable decoding?

Dr. J. 25 Aug 2026 · 12:22

Would love to see more transparency on how bias in training datasets might shape the decoded language-accuracy rates sound exciting, but could they lock users into limited expressions?

J.P.R. 3 25 Aug 2026 · 12:22

What guarantees the system’s adaptability to individual neurodiversity? If the training sets exclude certain firing patterns, won’t artificial speech still feel like an imposed language rather than the user’s own?

le_sceptique 25 Aug 2026 · 11:44

I’m skeptical about the long-term reliability of decoding motor-neuron patterns into coherent language. How do they handle natural language variability or emotions in speech?

Your Linux servers, as a desktop.
TermalOSSponsored
Ops, reimagined

Your Linux servers, as a desktop.

Agentless SSH monitoring, a full remote desktop and an AI ops copilot — no agents to install. Everything stays on your machine.

Get early access
主题
浏览
信息