JOVANA
Explore Library Glossary Getting Started Three Levels Fields How it works Mission
Join the mission
All guides

Balance, and Listening to the Listener

The vestibular prosthesis writes head motion back into the balance nerve; auditory attention decoding reads the brain to steer what the ear receives. Two frontiers, one loop.

Restoring the sense of motion

People with bilateral vestibular hypofunction lose the reflexes that stabilize gaze and posture; they suffer chronic imbalance and oscillopsia — the world bounces when they move. The vestibular prosthesis borrows the cochlear-implant playbook: MEMS gyroscopes sense head angular velocity, and that signal is encoded onto pulse trains delivered to the ampullary nerves of the three semicircular canals, aiming to restore the vestibulo-ocular reflex (VOR).

G = -\frac{\dot{\theta}_{\text{eye}}}{\dot{\theta}_{\text{head}}}, \qquad G \to 1

VOR gain: the eyes must counter-rotate at the head's velocity to hold an image still. A working prosthesis drives the measured gain back toward unity.

The baseline problem and canal crosstalk

There is a twist absent from the cochlea: vestibular afferents have a high spontaneous firing rate and encode direction by modulating up and down around that baseline. So the prosthesis must first establish a tonic baseline with constant stimulation, then modulate it — you cannot represent a leftward turn without room to fire below baseline.

r(t) = \operatorname{clip}\!\bigl(r_0 + k\,\dot{\theta}_{\text{head}}(t),\; 0,\; r_{\max}\bigr)

Rate-modulation encoding: head angular velocity shifts afferent firing above and below a tonic baseline r₀, saturating at the safe rate limits.

The read side: decoding who you are listening to

Everything above is writing. But the auditory frontier has a powerful read branch too. In a crowded room a listener attends to one talker; the cortex tracks that talker's speech envelope more strongly than the others'. Central auditory decoding — auditory attention decoding — reconstructs the attended envelope from neural activity, letting a device infer which voice you are trying to hear.

\hat{s}(t) = \sum_{\tau} g(\tau)\, r(t-\tau), \qquad \text{attended} = \arg\max_i\, \rho\bigl(\hat{s},\, s_i\bigr)

A backward (stimulus-reconstruction) decoder maps neural response r to an estimated envelope ŝ; the attended talker is the one whose real envelope correlates best with the reconstruction.

Close the loop and you get a cognitively-controlled hearing device: read the attended talker, then steer a beamformer to amplify exactly that voice. This is the read/write asymmetry turned to advantage — here reading is easier than writing, and it makes the write side (the hearing aid or implant) smarter. It is still a lab result: decoding needs seconds of data, degrades with more talkers, and robust real-time operation in daily life is unsolved. Related passive-attention ideas connect to the broader attention and workload BCI literature.