Implantable Neural Speech Decoders: Recent Advances, Future Challenges

The social life of locked-in syndrome (LIS) patients is significantly impacted by their difficulties to communicate. Consequently, researchers have started to explore how to decode intended speech from neural signals directly recorded from the cortex. The first studies in the late 2000s reported modest decoding accuracies. However, thanks to fast advances in machine learning, the most recent studies have reached decoding accuracies high enough to be optimistic about the clinical benefit of neural speech decoders in the near future. We first discuss the selection criteria for implanting a neural speech decoder in LIS patients, emphasizing the advantages and disadvantages associated with conditions such as brainstem stroke and amyotrophic lateral sclerosis. We examine the key design considerations for neural speech decoders, demonstrating how successful implantation requires careful optimization of multiple interrelated factors including language representation, cortical recording areas, neural features, training paradigms, and decoding algorithms. We then discuss current approaches and provide arguments for potential improvements in decoder design and implementation. Finally, we explore the crucial question of who should learn to use the neural speech decoder-the patient, the machine, or both. In conclusion, while neural speech decoders present promising avenues for improving communication for LIS patients, interdisciplinary efforts spanning neurorehabilitation, neuroscience, neuroengineering, and ethics are imperative to design future clinical trials.