Speech neuroprostheses hold the promise of reinstating communication abilities in individuals with paralysis, yet achieving naturalistic speed and expressiveness remains challenging (Moses, The New England Journal of Medicine 385:217, 2020). In our study (Metzger, Nature 620:1037, 2023), we utilized high-density surface recordings from the speech cortex of a participant in a clinical trial who had significant paralysis in limbs and voice. This enabled high-performance real-time decoding in three distinct speech-related outputs: text, speech audio, and facial-avatar animation. We trained deep-learning models on neural data gathered while the participant tried to silently articulate sentences. For text, our results showed high-speed, large-vocabulary decoding with a median of 78 words per minute and a 25% median word error rate. In terms of speech audio, we achieved intelligible speech synthesis that could be personalized to the participant’s pre-injury voice. For facial-avatar animation, we successfully decoded virtual movements of the mouth and face for both speech and non-verbal communication gestures. The decoders attained high efficiency with less than 2 weeks of training. Our research presents a novel multimodal speech-neuroprosthetic method that offers significant potential to restore comprehensive, embodied communication for individuals with severe paralysis.