The passive information would be useful if that would work with your inner voice.
If I find the article I will update.
also suggests a throat mic would do wonders
Is there some kind of sensor that one could wear that would be unobtrusive enough to measure subvocalization? If so, building a training dataset for an ML model would be as simple as having many people read and write a significant amount of text while using the sensor.
Even better if that text corpus overlaps with text that has been used to train text-to-speech models like Whisper, as you might get away with knowledge transfer with such a model.
It’s definitely worth looking into!
Isn't that what Musk's Neurolink is trying to achieve?