下一次当你看到软件工程师嘴唇微动却不出声时，可得三思而后行，别急着问他们是否需要请个“心理健康假” —— 他们可能只是戴上了一副智能眼镜，能在完全不发声的情况下与电脑对话。
The next time you see a software engineer shaping silent words with their mouth, you may have to think twice before asking if they need a mental health day – they might just be wearing a pair of smart glasses that let them talk to their computer without uttering a sound.
Subvocal 是一家获得 Y Combinator（YC）孵化支持的初创公司，致力于打造这类眼镜，且已有可工作的原型机，在 YouTube 视频中展示过，其中一段视频里 YC 合伙人 Dave Lieb 亲自试用，印象深刻到决定出资投资。公司指出，演示用的头戴装置并非最终版外观，这倒是好事：演示机明显是 3D 打印的，离“时尚”还差得远。根据官网展示的效果图，量产版会流畅得多。实际发布时间尚不确定，预订页面显示将于 2027 年第一季度推出，早鸟价 499 美元；最终正式售价目前也不详。
Subvocal is a Y Combinator-backed startup aiming to produce these glasses, and already has working prototypes to show off in YouTube videos, including one with YC partner Dave Lieb trying them out and being impressed enough to offer funding. The headset isn’t what the final version will look like, the company notes, which is good: The demo headset is obviously 3D printed and far from fashionable. Once the final product is available it should look a lot sleeker, based on mockups on the company’s website. It’s not clear when they’ll actually be released, though the preorder page says it’ll be sometime in Q1 2027 at an early bird price of $499. It’s not clear what the final full price will be.
关于工作原理，创始人、佐治亚理工学院计算机科学专业学生 Aarush Yadav 和 Arnav Kumar 在 YC 的发布帖中解释：设备从眼镜向佩戴者面部发射低功率无线电波；当佩戴者说话或仅仅张嘴默念单词时，面部肌肉运动和面部轮廓变化会扭曲这些波，眼镜捕捉到扭曲信号并将其转译为文字。
As for how the glasses work, the founders, Georgia Tech computer science students Aarush Yadav and Arnav Kumar, explain in a launch post on YC that the device uses low-powered radio waves projected from the glasses toward a wearer’s face. Those waves are distorted by muscle movements and the change in a person’s face shape as they speak or simply mouth words, which the duo said can be picked up and translated into written text.
“我们大约四个月前在一次黑客马拉松上做出第一版后开始着手研发，”两人在 YC 帖子中写道。“从那时起，我们已从一桌子电子元件和贴在脸上的天线，进化到半便携式电路板和外观流畅的眼镜，能在较小数据集规模下以 200 多词/分钟的速度解码开放式语音，且错误率相对较低。”
“We started working on this about four months ago after building the first version at a hackathon,” the pair explained in their YC post. “Since then, we’ve gone from a table full of electronics and antennas taped to our face to a semi-portable board and sleek glasses that can decode open-ended speech at 200+ WPM and a reasonably low error rate even at a small dataset scale.”
这仍属于概念验证阶段，许多关键细节尚未披露，例如准确率、续航时间、是否需配合 App 使用等。
This is just a proof of concept, and many key details are still missing, like the accuracy rate, battery life, whether it will paired with an app, or what-have-you.
“我们首先想要证明的核心问题是：信号是否存在，以及我们能否对其进行解码，”Yadav 和 Kumar 表示。在 Hacker News 上发布的宣布公司成立的帖子中，Yadav 和 Kumar 解释说，他们在黑客马拉松期间启动了该项目，重点是使用贴在 Yadav 脸上的 EMG（肌电图）电极。他们没有赢得黑客马拉松，但显然演示效果足够令人印象深刻，从而获得了旧金山风投公司 Afore Capital 的初始融资。
“The core thing we wanted to prove first was that the signal exists and that we can decode it,” Yadav and Kumar said. In a Hacker News post announcing the company, Yadav and Kumar explained that they started the project during the hackathon by focusing on using EMG electrodes attached to Yadav’s face. They didn’t win the hackathon, but apparently the demo was impressive enough to earn some initial funding from San Francisco VC firm Afore Capital.
不久后，两人从 EMG 转向了 RF（射频）感应，因为一款需要将电极贴在脸上的商业听写产品不太可能成功。“尽管此前已有学术研究，但要让这套系统跑通是一个巨大的工程挑战，”Yadav 说。“首先我们构建了一个极低噪声的 RF 感应系统。然后我们必须放置天线，使其场与发音器官最大程度重叠。我们在眼镜上的第一次尝试完全失败了，所以我们回归老习惯，把天线贴在脸上。”
Shortly after that, the pair pivoted from EMG to RF sensing, as a commercial dictation product that required sticking electrodes to one’s face was unlikely to succeed. “Despite prior academic research, getting this to work was a huge engineering challenge,” Yadav said. “First we built an extremely low-noise RF sensing stack. Then we had to place antennas so their fields overlapped maximally with the articulators. Our first attempt on glasses failed completely, so we reverted to old habits and taped antennas on our face.”
他们最终得到了现在的功能原型，并预训练了执行语音识别的 AI 模型，以尽量减少设置时间。据公司介绍，Lieb 只需朗读 20 个句子即可将头戴设备校准至其语音模式。“我们认为，无声语音可以成为与计算机交互的下一代界面，取代语音和键盘，”公司谈及其目标时表示。“邮件、提示词、消息、代码、笔记、指令，全都能以语速完成，且不打扰身边任何人。”
They eventually got to the functional prototype they have now, and pre-trained the AI models doing the speech recognition to minimize setup time. According to the company, Lieb only had to read 20 sentences to calibrate the headset to his speech patterns. “We think silent speech can become the next interface for interacting with computers, replacing voice and the keyboard,” the company says of its goal. “Emails, prompts, messages, code, notes, commands, all at the speed of speech without bothering anyone around you.”
凭借短短四个月的工作成果，Subvocal 可能会大获成功，为各行各业的工作者开启无声自语高效办公的新时代，也可能在试图规模化时折戟沉沙。有兴趣一探究竟的人现在可以预订一副，价格 199 美元，剩余 300 美元将在眼镜发货时支付。®
With just four months of work under their belts, Subvocal could be a massive hit, ushering in a new era of silent muttering productivity for workers in a variety of industries, or it could flop as it tries to scale. Those interested in finding out can preorder a pair now for $199, with the other $300 not due until the glasses ship. ®