This study demonstrates that internal speech can be decoded from single neurons in the supramarginal gyrus (SMG) of participants with tetraplegia. Two participants with implanted microelectrode arrays in the SMG and primary somatosensory cortex (S1) performed internal and vocalized speech of six words and two pseudowords. Significant neural representations of internal and vocalized speech were found in the SMG at both the single neuron and population levels. Offline decoding accuracy averaged 55% and 24% for each participant, while online decoding accuracy averaged 79% and 23%. Shared neural representations between internal speech, word reading, and vocalized speech were observed in participant 1, suggesting phonetic encoding in the SMG. S1 activity was modulated by vocalized but not internal speech, indicating no articulator movements during internal speech. The study provides a proof-of-concept for a high-performance internal speech brain–machine interface (BMI). Speech is a fundamental form of human communication, often lost in neurological diseases like amyotrophic lateral sclerosis (ALS). Brain–machine interfaces (BMIs) offer a promising solution by recording neural activity directly from the cortex. Cognitive BMIs have shown potential in restoring independence for tetraplegic patients by reading movement intent from the brain. Decoding internal speech signals could restore communication for those who have lost it. Decoding speech directly from the brain presents unique challenges. Non-invasive methods like fMRI, EEG, and MEG lack the necessary temporal and spatial resolution for online speech BMI. Intracortical electrophysiological recordings offer higher signal-to-noise ratios and better temporal resolution, making them suitable for internal speech decoding. Invasive methods like electrocorticography (ECoG) and stereo-electroencephalographic depth arrays have been used to decode vocalized and attempted speech. However, internal speech decoding remains challenging due to lack of behavioral output, lower signal-to-noise ratio, and differences in cortical activation compared to vocalized speech. Evidence suggests that the SMG is involved in subvocal speech and phonologic processing. Studies have shown that SMG activity is modulated during vocalized and internal speech, and that lesions in the SMG affect inner speech rhyming tasks. These findings support the possibility of an internal speech decoder from neural activity in the SMG. The relationship between inner speech and vocalized speech is still debated. While there is consensus on similarities between internal and vocalized speech processes, the degree of overlap is not well understood. Characterizing similarities between vocalized and internal speech could provide evidence that results from vocalized speech could translate to internal speech. The study found that the SMG represents words and pseudowords, providing evidence for phonetic encoding. The decoder achieved high classification with multiple internal speech strategies (auditory imagination/visual imagination). Activity in S1 was modulated by vocalized but not internal speech, suggestingThis study demonstrates that internal speech can be decoded from single neurons in the supramarginal gyrus (SMG) of participants with tetraplegia. Two participants with implanted microelectrode arrays in the SMG and primary somatosensory cortex (S1) performed internal and vocalized speech of six words and two pseudowords. Significant neural representations of internal and vocalized speech were found in the SMG at both the single neuron and population levels. Offline decoding accuracy averaged 55% and 24% for each participant, while online decoding accuracy averaged 79% and 23%. Shared neural representations between internal speech, word reading, and vocalized speech were observed in participant 1, suggesting phonetic encoding in the SMG. S1 activity was modulated by vocalized but not internal speech, indicating no articulator movements during internal speech. The study provides a proof-of-concept for a high-performance internal speech brain–machine interface (BMI). Speech is a fundamental form of human communication, often lost in neurological diseases like amyotrophic lateral sclerosis (ALS). Brain–machine interfaces (BMIs) offer a promising solution by recording neural activity directly from the cortex. Cognitive BMIs have shown potential in restoring independence for tetraplegic patients by reading movement intent from the brain. Decoding internal speech signals could restore communication for those who have lost it. Decoding speech directly from the brain presents unique challenges. Non-invasive methods like fMRI, EEG, and MEG lack the necessary temporal and spatial resolution for online speech BMI. Intracortical electrophysiological recordings offer higher signal-to-noise ratios and better temporal resolution, making them suitable for internal speech decoding. Invasive methods like electrocorticography (ECoG) and stereo-electroencephalographic depth arrays have been used to decode vocalized and attempted speech. However, internal speech decoding remains challenging due to lack of behavioral output, lower signal-to-noise ratio, and differences in cortical activation compared to vocalized speech. Evidence suggests that the SMG is involved in subvocal speech and phonologic processing. Studies have shown that SMG activity is modulated during vocalized and internal speech, and that lesions in the SMG affect inner speech rhyming tasks. These findings support the possibility of an internal speech decoder from neural activity in the SMG. The relationship between inner speech and vocalized speech is still debated. While there is consensus on similarities between internal and vocalized speech processes, the degree of overlap is not well understood. Characterizing similarities between vocalized and internal speech could provide evidence that results from vocalized speech could translate to internal speech. The study found that the SMG represents words and pseudowords, providing evidence for phonetic encoding. The decoder achieved high classification with multiple internal speech strategies (auditory imagination/visual imagination). Activity in S1 was modulated by vocalized but not internal speech, suggesting