Show HN: Subvocal – Talk to your computer without making noise
Subvocal introduces groundbreaking smart glasses that translate silent mouthing into text using advanced RF technology and AI, promising a discreet, futuristic interface for interacting with computers. This 'Show HN' captivated the community with its technical depth, addressing common voice input frustrations while sparking debates on privacy, practicality, and sci-fi visions.
The Lowdown
Subvocal, a startup from YC W26, is developing smart glasses that allow users to communicate with their computers by silently mouthing words, which are then decoded into text at speech speed. This innovative approach aims to overcome the social awkwardness and privacy concerns associated with audible voice commands, leveraging a deep dive into silent speech technology.
The journey began at a hackathon, where founders Aarush and Arnav, without prior experience in silent speech, recreated MIT's AlterEgo using EMG electrodes. This initial success, despite its crudeness, solidified their vision.
- Technical Pivot: They abandoned EMG due to practical limitations (sweat, electrical interference, placement variability, user discomfort) and transitioned to RF technology.
- RF Mechanism: The glasses emit low-power radio waves that pass through the face. These waves are modulated by the subtle movements of the tongue, lips, and soft palate during silent speech. Proprietary foundation models then decode these RF perturbations into text.
The Gossip
Form Factor Fusses and Future Formats
Many commenters, while impressed by the technology, questioned the practicalities of the glasses form factor. Concerns included compatibility with existing prescription eyewear, the social perception of silently mouthing words in public, and preferences for alternative designs like earbuds, necklaces, or even a laptop-mounted device. The founders clarified that the RF technology's current limitations necessitate the glasses for optimal signal capture but are exploring other form factors for the future, while also explaining why lip-reading from cameras or mmWave radar proved insufficient.
Privacy Perils and Processing Predicaments
A significant thread of discussion revolved around data privacy and the implications of silent speech being processed by a third party. Users expressed strong concerns about their 'internal thoughts' being sent to the cloud, asking for offline capabilities and clarification on data usage. The founders acknowledged privacy as important and confirmed that, for now, decoding occurs via a companion app in the cloud, impacting offline functionality.
Speedy Speech vs. Typed Thoughts
The assertion that talking is faster than typing sparked debate. Many agreed, especially for fluid LLM prompting, noting speech's natural flow and higher words-per-minute (WPM) rates compared to typing. Others, however, felt they typed faster or more precisely, especially for structured tasks like coding, and found spoken language often contained more filler. The technology's futuristic nature also drew comparisons to sci-fi tropes, from 'Ender's Game' to '1984', highlighting both its potential and dystopian undertones.
Website Woes and Skepticism Surges
Several commenters criticized the product's website for its confusing user experience, lack of immediate clarity on the product, and insufficient technical details on the 'how it works' page. The use of AI-generated images for the product presentation also fueled skepticism among some, with accusations of a 'scam' and doubts about the feasibility of delivering a polished product by the ambitious Q1 2027 timeline. The founders responded positively to feedback, promising website improvements.
Technical Triumphs and Trialed Alternatives
The technical details of Subvocal's RF-based silent speech decoding were a major point of interest. The founders elaborated on their rigorous process, including the rejection of EMG, IMUs, and optical lip-reading (due to inherent information-theoretic limits like insufficient visemes) and mmWave radar. They explained the physics of RF modulation by facial tissue and how they overcame engineering challenges like low-noise sensing and cross-user generalization through extensive data collection.