Declarative Framing: What I Want Subvocal-Input-for-Agents to Look Like

A paradigm shift from custom neckbands to open-source middleware rails.

The Bottleneck: Hardware-First Myopia

For years, the silent speech interface (SSI) community has been stuck. The default approach—epitomized by projects like MindOS—is vertically integrated, custom-hardware myopic: tape a couple of gel electrodes to your jaw, hook them up to a microcontroller, train a Random Forest on 8 closed commands, and call a cloud LLM to run browser tasks.

This vertical hardware approach hits three walls: 1. The Vocabulary Ceiling: Whole-word classifiers max out at 5–8 commands before accuracy plummets. You can't browse the web if your jaw can only say "OPEN", "SEARCH", "CLICK", "SCROLL", "ENTER", and "CANCEL". 2. The Repositioning Cliff: Surface EMG is notoriously unstable. Reposition the device by 2 millimeters and accuracy drops by 15%. 3. The Capital Barrier: Manufacturing wearable neckbands requires capital, supply chains, and regulatory clearances that slow software builders to a crawl.

The Framing Shift: We Are the Rails, Not the Train

We are declaring a pivot. We are not building a proprietary neckband. We are building the open-source SDK that every subvocal and biosignal interface will use to control LLM-driven agents.

Whether you are hacking in your apartment with a $200 OpenBCI board, researching laryngeal activations in a university clinical lab with a Delsys Trigno rig, or designing Meta's next-generation wrist sEMG controller—you need the same connective software rails. You shouldn't have to rewrite filters, feature extractors, noise simulators, and context-aware decoders from scratch.

By providing a unified developer platform, we decouple the sensor hardware from the cognitive agent.

What We Are Shipping Today (v0.1.0-alpha)

We are releasing the core building blocks of the Subvocal SDK:

1. The Compressed Shorthand Layer

To bypass the vocabulary ceiling, we introduced a spelling-shortening grammar. Users "silent-speak" compressed consonant shorthand (e.g. g gl for Google). We model sEMG signal degradation through an articulatory phonetic noise simulator, mapping jaw/throat activations into 5 physiological sEMG clusters.

2. Context-Aware Intent Reconstruction

Our decoder resolves noisy shorthand queries using: - Asymmetric sEMG Levenshtein Distance: A dynamic alignment matrix that highly discounts consonant omissions (which are common when users drop vowels/consonants in silent speech) while penalizing insertions. - Command-Aware Prioritization: It filters and ranks candidate targets (URLs, contacts, calendar events, active UI elements) based on the action intent (e.g. a CLICK query only searches active screen elements; a TYPE query only searches user contacts).

3. Modular ML Core (emg_core)

We have built and verified the physiological signal pipeline: - Reconciled DSP Filters: Defaulting to the AlterEgo-inspired 1.3–50.0 Hz bandpass filter (physiologically correct for slow articulatory neuromuscular activations) with built-in options to support standard 20.0–450.0 Hz gestures. - Deep Learning Skeletons: Local training pipelines for standard Random Forest, 1D CNN (PyTorch), and GRU (PyTorch) classifiers. - Model Serialization: Saving/loading parameters, feature scalers, and metadata to .pth and .joblib objects.

4. Zero-Dependency TTS Audio Feedback

An offline text-to-speech engine that automatically falls back to macOS native say and afplay utilities, providing sub-millisecond audio confirmation without introducing bloated python dependencies.

The Roadmap Ahead

We are standardizing this interface. In the coming weeks, we will expose these subvocal commands as a Model Context Protocol (MCP) server, making it trivial to plug a biosignal stream into Claude Desktop or any agentic system.

The code is fully open-source. Clone it, run the integration tests, and start building.

Join the conversation on GitHub.