the problem:

roughly 1.4 billion people globally have some form of speech or language impairment. existing assistive communication tools — gaze trackers, button grids, eye-scanning systems — are slow, exhausting, and far behind the pace of natural thought. they translate intent through a degraded physical channel. vocl takes a different approach: instead of routing around the body, it reads the intent before it fully becomes physical.

how it works:

EMG sensors placed on the skin surface capture the electrical impulses produced by subvocal muscle movement — the micro-activations that occur even when you try to speak without making sound. a signal processing pipeline filters noise and segments the data into time windows. a neural network, trained on labeled EMG recordings, maps these patterns to phonemes and words. the output is passed through a text-to-speech engine to produce audible speech in near real time.

technical approach:

the core model is a CNN-LSTM architecture suited for temporal EMG classification. training data spans both voiced and subvocalized speech attempts. the inference pipeline is optimized for low latency — the gap between thought and audible output needs to feel like speech, not like typing.

who it's for:

ALS, laryngeal cancer, locked-in syndrome, severe dysarthria. anyone whose voice has been taken from them, or who was never given one. vocl is built to close that gap.