To address this, we propose two simple yet effective strategies—a new way to train an attempted speech BCI to ignore inner speech, and a password-protection system that prevents inner speech from being decoded unless the user first imagines the password

But there’s a challenge: if inner speech lives in motor cortex, could a BCI accidentally decode something a user only intended to think, not say? In controlled cognitive tasks, like silently counting shapes in a grid, we found BCIs could indeed pick up number words.

Inner speech was highly correlated with attempted speech. But we discovered a unique “motor-intent” dimension that separates them—with ALS who was completely anarthric—suggesting attempted speech remains a viable BCI driver even in near locked-in states.

For people who’ve lost the ability to speak, a BCI decoding inner speech might be more comfortable by bypassing physical effort while still enabling fluent communication. Here T16 uses an attempted speech BCI :speaking:

We investigated inner speech in the motor cortex—finding a robust representation that enabled real-time decoding of imagined sentences with error rates as low as 14% for a 50-word and 26% for a 125,000-word vocabulary

Indeed, typical speech decoders inadvertently decode inner speech with very high frequency (blue). However, an “imagery silenced” decoder training strategy was able to prevent this from happening with high accuracy (gray).

Word representations were highly correlated between overt speech (solid) and inner speech (dashed), and even listening and reading! This raises a risk of unintentional decoding; however, behaviors were still separable in neural state space along an “overtness” axis.

We show that internally spoken sentences can be decoded in real-time with accuracies between 14-33% error rate for a 50 word vocabulary. This suggests that verbal thought is strongly encoded in areas of motor cortex typically targeted for speech BCIs.

First post! Such an awesome experience working with the rest of the Stanford team and participant T12