The Subvocal Company

The Subvocal Company is a San Francisco hardware startup building Subvocal, a pair of glasses meant to turn silently mouthed words into text on a computer. The wearer articulates words without making a sound, sensors in the frame pick up the small facial movements involved, and the company's models decode those signals into text that the company intends to work in any application where the user would normally type ^[1][2][3]. The company was founded in 2026 by two Georgia Institute of Technology (Georgia Tech) computer science students, Aarush Yadav (CEO) and Arnav Kumar (CTO), and is part of Y Combinator's Fall 2026 batch ^[1]. It reported about $1.3 million in pre-seed funding from Afore Capital, Y Combinator and angel investors ^[6][7], and it opened paid reservations for the glasses on 30 October 2026, with shipping estimated by the company for the first quarter of 2027 ^[2][4].

Subvocal belongs to a long line of research on silent speech interfaces, which try to recover words from the movements or muscle signals of speech rather than from sound. Earlier systems such as the MIT Media Lab's AlterEgo (2018) used electrodes on the face and jaw, while Cornell's EchoSpeech glasses (2023) used sonar ^[10][11]. Subvocal's launch materials describe a non-contact radio-frequency (RF) approach instead ^[2][5].

Overview

History

TreeHacks prototype

Yadav and Kumar were computer science students at Georgia Tech when they attended TreeHacks, Stanford University's annual hackathon, in February 2026 ^[6][8]. According to IndianWeb2, they worked in a four-person team that built an early prototype using muscle-sensing electrodes, machine learning and language-model assistance to translate facial muscle signals into text and computer commands. The project did not win, but the founders decided the idea had consumer potential, raised money and moved to San Francisco to work on it full time ^[6][7].

Yadav gave his own account of the hackathon in "How we made Subvocal", an X Article published on 31 October 2026, the day after the public launch ^[16]. He wrote that they stuck EMG (electromyography) sensors to his face, recorded him subvocalizing the 26 letters of the alphabet, and within 24 hours trained a 26-class letter classifier that reached 80% accuracy, enough for him to prompt an AI system letter by letter ^[16]. Yadav's article dates the trip to TreeHacks to March, but the 2026 edition of the hackathon ran from the evening of Friday 13 February to the morning of Sunday 15 February 2026, according to The Stanford Daily, which matches the February date given by IndianWeb2 and the Georgia Tech profile ^[6][8][15].

A Georgia Tech John H. Martinson Honors Program profile of Yadav, published on 29 September 2026, says he and his cofounder began fundraising in late February and closed a $500,000 pre-seed round from Afore Capital, a pre-seed venture firm, within two to three weeks ^[8]. IndianWeb2 dates Afore's initial $500,000 investment to April 2026, when the product was still an early prototype, and the VCBacked funding database lists a $500,000 pre-seed round dated 6 April 2026 with Afore Capital as the investor ^[6][9]. The same Georgia Tech profile described the device at that stage as a wearable that "sits behind your ear" and decodes "subvocalized thoughts" into text from bio signals around the face, using what it called a "thought-to-text" AI model, and said the company then had "a working letter-level prototype that can spell out words using your thoughts at 30 words per minute" ^[8]. It also reported that Yadav planned to take a leave of absence from college with his cofounder to work on the company full time ^[8].

The company's own Launch YC post, published on 30 October 2026, gives a shorter timeline: it says the team "started working on this about four months ago after building the first version at a hackathon," and describes progress "from a table full of electronics and antennas taped to our face to a semi-portable board and sleek glasses" ^[2].

Y Combinator and funding

The Subvocal Company is listed in Y Combinator's Fall 2026 batch, with David Lieb as its primary group partner ^[1]. IndianWeb2 and Dealroom reported that the company won YC's Group 1 Product Showcase, which is voted on by YC alumni and batchmates, and the launch post says "a more crude version" of the product won the showcase a few weeks before launch ^[2][6][7].

On 30 September 2026, IndianWeb2 and Dealroom reported that the company had raised approximately $1.3 million in pre-seed backing. Both outlets describe the total as Afore Capital's initial $500,000, a follow-on investment from Afore, angel money, and a $500,000 investment commitment from Y Combinator ^[6][7]. In his X Article, Yadav wrote that "Gaurav over at" Afore backed the team "super early, right after seeing our initial makeshift EMG rig", before it had glasses or a refined hardware demo, and that David Lieb and Max Kolysh of Y Combinator later "saw the exact same promise" when the founders pitched YC ^[16]. Afore's website lists Gaurav Jain as a co-founder and managing partner of the firm ^[17]. Yadav was quoted as saying, "Computers have become far more powerful, but the way we talk to them has changed very little. We believe the best way to interact with computers and AI agents will be silent, seamless and far faster than what exists today" ^[6].

Public launch and preorders

The company launched publicly on 30 October 2026 with a Launch YC post and an X post by Yadav that said, "We've raised $1.3M from YC & Afore Capital so you never have to type or speak again," and announced that preorders were live ^[2][5]. Yadav's post had been viewed about 10.6 million times according to X's view counter as of 1 November 2026 ^[5]. The next day he published the "How we made Subvocal" article on the company's development ^[16].

Reception

Digital Trends, covering the launch on 30 October 2026, noted that the glasses were still in development and that the 200 words per minute figure had not been independently verified, and asked how they would handle accents, unfamiliar words and long dictation ^[22]. Startup Fortune reported that replies on X ranged from congratulations to doubts that mouthing words in public is less awkward than talking ^[23]. Within a day, developer Joe Blau published Subvox, an open-source macOS prototype that matches AirPods Pro head-motion data against a few user-trained words; its documentation describes a small-vocabulary template recognizer with no language model ^[23][24].

Product

How it works

The company describes Subvocal as an input method rather than a standalone app. According to the launch post, the glasses use RF sensing "to pick up the tiny facial movements you make while silently articulating words": they send low-power radio waves through the face, measure how the signal changes as the muscles move, and pass the result to models that decode it "directly into text" ^[2]. Yadav's announcement describes the same idea as tracking "your internal speech muscles with non-contact sensing" ^[5]. The company says the system uses no cameras and does no camera-based lip reading ^[3][4].

The site says the user "mouth[s] words as subtly as you want, without producing audible speech," and that while someone up close may notice slight movement, the glasses are designed to be invisible at normal conversational distance ^[3]. Because it does not rely on sound, the company argues that background noise does not interfere the way it does with voice dictation ^[3].

The FAQ also sets limits on what the device does. It states that it "does not read arbitrary or involuntary thoughts" and "only decodes words you intentionally articulate," and that it is not always listening: the user chooses when to articulate input ^[3]. Hiring listings on its YC page include roles for a Head of ML, a Head of Sensing and a Head of Hardware ^[1].

Choice of sensing method

In his 31 October article, Yadav wrote that the team shortlisted and tried four sensing approaches before choosing RF. The assessments below are his; the company has not published the underlying experiments ^[16]:

Yadav also wrote that visual speech recognition work "by the likes of Meta has hit a floor at ~20% WER" ^[16]. Published results on the LRS3 lip-reading benchmark do not show a fixed floor. Meta AI's AV-HuBERT reported a 26.9% word error rate in 2022 ^[19]; Auto-AVSR, from Imperial College London and Meta AI researchers, reported 19.1% in 2023 ^[20]; SynthVSR, by researchers mostly at Meta AI, reported 16.9% in 2023 using synthetic lip movements and publicly available data ^[21]; and Google researchers reported 12.8% the same year with a Conformer-based model trained on about 100,000 hours of transcribed YouTube video ^[25]. These TED-talk benchmark figures are not directly comparable with the team's webcam test.

For the RF approach, Yadav wrote that the company built on Wagner et al. (2022), which he described as the work that first came up with the idea of through-body RF sensing for silent speech recognition and proved that the mechanism works ^[16]. That paper, by Christoph Wagner, Petr Schaffer and colleagues at Technische Universität Dresden and published in Scientific Reports in March 2022, used custom stepped frequency continuous wave radar to measure how signals passing between three antennas taped to the cheeks and chin changed during speech, 100 times a second over 1 to 6 GHz ^[18]. On a 50-word German command corpus recorded from two male speakers over three sessions each, it reported 99.17% speaker-dependent multi-session accuracy and 88.87% inter-session accuracy, where the antennas were removed and reattached between sessions ^[18]. The paper itself calls radar-based silent speech an idea "around for over 20 years" and builds on a 2018 study by Peter Birkholz and colleagues, two of whom are also co-authors of the 2022 paper, showing that the reflection and transmission spectra of antennas placed on the facial skin carry information about vocal tract shape. Its speakers said the words aloud during recording, and the task was isolated-word classification, not continuous speech ^[18].

Claimed performance

The company claims its glasses decode open-ended silent speech at more than 200 words per minute ^[2][3][5]. The launch post adds that this works "at a reasonably low error rate even at a small dataset scale," but the launch materials give no error rate or test protocol ^[2][3]. In his 31 October article, Yadav wrote that the current prototype glasses "can decode continuous speech on a new user after 1 minute of calibration with 85-92% word accuracy"; he did not describe the vocabulary, number of test users or evaluation protocol behind that figure ^[16]. The FAQ says accuracy "continues to improve as we collect more data and train larger models," that the system "can occasionally get something wrong," and that the product relies on context and easy corrections ^[3]. The 200-words-per-minute figure is self-reported; it compares with the 30 words per minute letter-level prototype described in the Georgia Tech profile ^[8].

According to the FAQ, there will be "a short personal calibration process" when a user first sets up the glasses ^[3]. The launch post says the company plans to collect thousands of hours of "population-scale RF data" over the following months to pretrain a model that needs only very short calibration for a new wearer ^[2]. It also stated that the prototype hardware "looks a lot worse than the product we intend to ship," that the website shows the intended final design, and that the team expected to miniaturize the electronics enough to fit entirely on the glasses within about a week of launch ^[2].

Intended uses

The launch post frames Subvocal around voice input in places where speaking aloud is awkward, such as open offices, cafes and public transit, and names heavy users of Wispr Flow, Superwhisper, ChatGPT voice and system dictation as its first target customers ^[2]. The founders argue that as AI agents take on more work, prompting becomes a bigger share of computer use, which favors fast, open-ended input ^[2]. The company's site shows a user dictating a prompt to Claude Code next to coworkers, and its FAQ lists ChatGPT, Claude, Cursor, Slack, email and documents as places it should work ^[3].

Preorder terms

As of 1 November 2026, the preorder page lists the following terms ^[4]:

The company says the product images show its "current design direction" and that final specifications are still being developed ^[4]. Dealroom's 30 October update gave the estimated shipping window as the second quarter of 2027 and described a first batch of 100 pairs; the company's own preorder page says Q1 2027 and does not state a batch size ^[4][7].

Founders

Background: silent speech interfaces

A silent speech interface recovers words from the articulation of speech, the movements of the lips, jaw and tongue or the muscle activity behind them, rather than from audio. Unlike conventional speech recognition, it can in principle work in noisy places, keep input private, and give a voice to people who cannot vocalize, uses that researchers behind AlterEgo and EchoSpeech have pointed to ^[10][11]. Unlike a brain-computer interface, it reads deliberate articulation rather than brain activity. Researchers have used several sensing methods:

AlterEgo was led by Arnav Kapur in Pattie Maes's Fluid Interfaces group and presented at the ACM Intelligent User Interfaces conference. Its network was trained to match signals to particular words and could be customized to a new user by retraining its last two layers ^[10]. EchoSpeech's authors argued that acoustic sensing avoids the privacy and battery problems of wearable cameras, reporting about ten hours of battery life with acoustic sensing against 30 minutes with a camera ^[11]. Radar researchers have made a similar case for their approach: Lee and colleagues note that many silent speech methods need invasive procedures or sensors taped or glued to the skin, while IR-UWB radar works without contact and is robust to outside light and sound ^[13].

In February 2026, UploadVR reported that Apple had acquired Q.ai, a startup working on silent speech, reportedly for about $1.6 billion; Q.ai's patents point to an optical approach that senses light reflected from the face ^[14]. The same report noted that a startup called AlterEgo was pursuing a jaw-sensor approach, and that Meta was instead betting on handwriting sensed by a wrist-worn sEMG band (the Meta Neural Band) ^[14]. Subvocal's eyeglasses form factor also places it alongside the wider category of AI glasses; UploadVR argued that private text entry in public will be crucial if glasses are ever to supplant the smartphone ^[14].

Continue reading

• Y CombinatorY Combinator (YC) is an American startup accelerator and early-stage venture firm headquartered in San Francisco.Mentioned 8 times above
• SSpeech RecognitionSpeech recognition, usually called automatic speech recognition (ASR), is the computational task of converting a spoken-language signal into a sequence of written symbols.Mentioned 4 times above
• AAI GlassesAI glasses are eyeglasses that carry a camera, microphones, open-ear speakers and an AI assistant, and in most cases no display at all.Listed in See also
• BBrain-Computer InterfaceA brain-computer interface (BCI), also called a brain-machine interface, is a system that measures activity in the nervous system and translates it into commands for an external device, bypassing muscles and…Listed in See also
• Meta Neural BandThe Meta Neural Band is a wrist-worn input device that reads surface electromyography (sEMG) signals from the forearm and converts them into commands for Meta Ray-Ban Display glasses.Listed in See also
• Wispr (Wispr Flow)Wispr (legally Wispr AI, Inc.) is a San Francisco artificial intelligence company best known for Wispr Flow, an AI voice dictation application that turns natural speech into clean, formatted text inside any…Listed in See also

More from See also

• Conduit (neurotechnology company)

More to read

• Meta AI
• Word error rate
• Superwhisper
• Google
• Apple Inc.

Topics:AI CompaniesAI HardwareConsumer AI DevicesSpeech & Audio AI

References

1. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12Y Combinator. "The Subvocal Company: Talk to your computer without talking". YC company directory. Accessed 1 November 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15 ^16Yadav, Aarush. "The Subvocal Company: Talk without talking". Launch YC, 30 October 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12The Subvocal Company. "Subvocal". Company website. Accessed 1 November 2026.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7The Subvocal Company. "Preorder Subvocal". Preorder page. Accessed 1 November 2026.
5. ^1 ^2 ^3 ^4 ^5 ^6Yadav, Aarush (@AarushYadav47). Post announcing the $1.3M raise and preorders. X, 30 October 2026.
6. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11IndianWeb2 Desk. "Subvocal's WearableTech Raises $1.3M from Y Combinator & Others to Transform Silent Speech Into Seamless Computer Interaction". IndianWeb2, 30 September 2026.
7. ^1 ^2 ^3 ^4 ^5 ^6Dealroom. "Subvocal raises $1.3M pre-seed to turn silent speech into computer commands". Dealroom News, 30 September 2026 (updated 30 October 2026).
8. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8Smith, Jalana. "JMHP Student Secures $500k Seed Funding for Startup". John H. Martinson Honors Program, Georgia Institute of Technology, 29 September 2026.
9. ^VCBacked. "The Subvocal Company". VCBacked company directory. Accessed 1 November 2026.
10. ^1 ^2 ^3 ^4Hardesty, Larry. "Computer system transcribes words users 'speak silently'". MIT News, 4 April 2018.
11. ^1 ^2 ^3 ^4DiPietro, Louis. "AI-equipped eyeglasses can read silent speech". Cornell Chronicle, 6 April 2023.
12. ^Shin, Young Hoon; Seo, Jiwon. "Towards Contactless Silent Speech Recognition Based on Detection of Active and Visible Articulators Using IR-UWB Radar". Sensors 16(11):1812, 29 October 2016.
13. ^1 ^2Lee, Sunghwa; et al. "IR-UWB Radar-Based Contactless Silent Speech Recognition of Vowels, Consonants, Words, and Phrases". arXiv:2312.09572; IEEE Access, 2023.
14. ^1 ^2 ^3Heaney, David. "Apple Acquired 'Silent Speech' Startup Q.ai". UploadVR, 2 February 2026.
15. ^Williams, Sofia. "12th annual TreeHacks awards $500,000 in prizes". The Stanford Daily, 15 February 2026.
16. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9Yadav, Aarush. "How we made Subvocal". X Article, 31 October 2026 (shared in this post).
17. ^Afore Capital. "Afore Capital". Company website. Accessed 1 November 2026.
18. ^1 ^2 ^3 ^4Wagner, Christoph; Schaffer, Petr; Amini Digehsara, Pouriya; Bärhold, Michael; Plettemeier, Dirk; Birkholz, Peter. "Silent speech command word recognition using stepped frequency continuous wave radar". Scientific Reports 12, 4192, 9 March 2022.
19. ^Shi, Bowen; Hsu, Wei-Ning; Lakhotia, Kushal; Mohamed, Abdelrahman. "Learning Audio-Visual Speech Representation by Masked Multimodal Cluster Prediction". arXiv:2201.02184, January 2022.
20. ^Ma, Pingchuan; et al. "Auto-AVSR: Audio-Visual Speech Recognition with Automatic Labels". arXiv:2303.14307, March 2023.
21. ^Liu, Xubo; et al. "SynthVSR: Scaling Up Visual Speech Recognition With Synthetic Supervision". arXiv:2303.17200, March 2023.
22. ^Mangalam, Sudhanshu Kumar. "Subvocal's glasses promise to let you talk to your computer in complete silence". Digital Trends, 30 October 2026.
23. ^1 ^2Haums, Mervik. "A Startup Raised $1.3 Million for Silent Speech Glasses and a Developer Built the Free Version in Nine Minutes". Startup Fortune, 31 October 2026.
24. ^Blau, Joe. "joeblau/subvox". GitHub repository, created 31 October 2026. Accessed 1 November 2026.
25. ^Chang, Oscar; Liao, Hank; Serdyuk, Dmitriy; Shah, Ankit; Siohan, Olivier. "Conformers are All You Need for Visual Speech Recognition". arXiv:2302.10915, February 2023.

Show all 25 references