Fig. 1: Extraction of the three speech representations.

A The dataset contains synchronized sEEG and audio data during speech production. Highlighted is one single channel’s neural signal (also in orange) and the audio signal of one word. B The articulatory trajectories (N = 20) were estimated using an acoustic-to-articulatory inversion (AAI) LSTM neural network. C The acoustic representation in the form of mel-spectrograms was calculated using short-time Fourier transform (STFT) with N = 23 filterbanks. D The semantic embeddings (160 dimensional) were extracted for each word using a word2vec model and a linear regression was subsequently trained to predict the neural timeseries. All data presented in this figure were extracted from sub-06.