subvocal

Physiological Silent Speech Interface Middleware

An open-source, hardware-agnostic SDK connecting surface electromyography interfaces to LLM-driven AI agents via compressed articulatory shorthand decoders.

What is Subvocal SDK?

Subvocal SDK is a library for sEMG bio-signal capture, digital signal processing, gesture classification, and context-aware natural language intent reconstruction.

[*]
Articulatory Shorthand: Bypasses the typical 5-8 whole-word vocabulary limit of standard sEMG classifiers.
[*]
Asymmetric Levenshtein: Aligns compressed phonetic shorthand inputs using custom physiological confusion clusters.
[*]
Context Prioritization: Dynamically filters and ranks candidate targets based on focused application states.
[*]
Signal Preprocessing: AlterEgo-inspired 1.3-50.0 Hz bandpass filtering defaults optimized for slow gestures.
[*]
Classifier Skeletons: Train Random Forest, 1D CNN, GRU, and Transformer model topologies locally on raw traces.
[*]
MCP Server: Exposes intent-reconstruction decoders as standard Model Context Protocol tools.
[*]
Offline TTS: Zero-dependency local audio feedback engine prioritizing native macOS audio tools.

Read Docs

Pipeline Performance & Baselines

[*]

Evaluated across 50 realistic shorthand command scenarios, the Subvocal SDK delivers 74.0% baseline heuristic reconstruction accuracy (approaching 100% on specific control groups) at <0.72 ms execution latency.

Built for privacy first

[*]

The Subvocal SDK operates securely and fully offline. Your biopotential raw metrics, feature vectors, and spelling shorthand logs are processed on local CPU threads, ensuring physiological metrics never leak to external cloud servers.

FAQ

Subvocal SDK is an open-source physiological silent speech middleware library. It captures muscle biopotentials, filters signal velocities, and reconstructs intents to drive local AI agents hands-free.
Rather than classifying whole words (which falls off in accuracy at ~8 words), users speak compressed phonetic shorthand (consonant strings). The SDK translates these strings using Levenshtein distance mapped to anatomical muscle groups.
No. While it supports OpenBCI Cyton and Delsys Trigno hardware controllers, the SDK includes a built-in mock channel noise simulator to test client intent decoders fully offline.
The SDK provides adapters for OpenAI GPT-4o, Anthropic Claude, Google Gemini, and local models running via Ollama.
We package our phonetic reconstruction decoders as stdio-based MCP tool servers. This enables compatible clients (like Claude Desktop) to parse and invoke subvocal inputs as standard tool calls.
Yes, the Subvocal SDK is fully open-sourced under the MIT License. You can use, modify, or extend it for academic research or commercial integrations.