Speech-enabling brain implants pass milestones

Brain-to-text decoding was achieved by the combination of two systems: a recurrent neural network (RNN, a type of artificial neural network), which ran algorithms that decipher brain activity associated with movements of articulators (parts of the vocal tract); followed by a language model that composed sentences at a rate of 78 words per minute (albeit with a 25.5% word error rate) from a set of 1,024 words. Importantly, the authors observed that neural activity recorded from a brain region (called Broca's area) widely thought to be crucial for speech production could not be decoded - raising questions about whether this area contains useful information for speech decoding. [...]it remains to be seen which BCI approach - MEAs or ECoG - will best serve the needs of users in terms of safety and long-term efficacy in real-life applications. MEAs capture rich functional information from a small cortical area, but the signals tend to be unstable and require frequent updating of speech-decoding models. [...]the longevity of MEAs might be limited by degradation of the electrode materials and tissue encapsulation of the devices15.