vocl

A non-invasive, AI-powered headset that restores speech for people with speech disorders by converting attempted speech into audible voice in real-time.

The Problem

7.5 million Americans with dysarthria, aphasia, paralysis, ALS, or stroke can't speak, but traditional solutions fail:

AAC tablets: $7,000-15,000, slow (10-15 words/min), requires typing

Invasive BCIs: $50,000-100,000+, requires brain surgery

Result: 30-50% abandonment rate, 70% unemployment, severe depression

How It Works

8 EMG electrodes on face/jaw capture electrical signals when user attempts to speak (even silently)

Machine learning classifies 44 English phonemes from muscle signals (82% accuracy)

LLM converts phoneme stream into natural language (reduces error rate from 28% to 12.2%)

Text-to-speech outputs voice in <300ms - fast enough for real conversation

Key Innovation

"Attempted speech" (not imagined) - captures 3-5× stronger signals when users TRY to speak

EMG (not EEG) - better spatial resolution, less noise, targets exact speech muscles

Non-invasive - no surgery required

Affordable - $2,500 target price (75% cheaper than AAC tablets, 95% cheaper than invasive BCIs)

Fast - 30-50 words/min vs. 10-15 WPM for typing-based AAC

Performance

82% phoneme accuracy (up from 68% after expert optimization)

<300ms latency (real-time conversation)

8-channel signal acquisition at 250Hz

Working prototype complete