Speech is the fastest and most natural form of communication, which can be impaired in certain disorders. Speech brain-computer interfaces (BCIs) offer a solution by decoding brain activity into speech. Current neuroprosthetic devices focus on the motor cortex, which might not be usable in all patient populations. Fortunately, many other brain regions have been associated with the speech production process. Here, we investigate which regions are potential (alternative) targets for a speech BCI across a brain-wide distribution within a single study. The distribution includes sulci and subcortical areas, sampled with both a high temporal and a high spatial resolution. Thirty participants were recorded with intracranial electroencephalography during speech production, resulting in 3249 recorded contacts across the brain. We trained machine learning models to continuously predict speech from a brain-wide global to a single-channel local scale. Within each scale we examined a variation of selected electrode contacts based on anatomical features within participants. We found significant speech detection in both gray and white matter tissue, no significant difference between gyri and sulci at any of the analysis scales and limited contribution from subcortical areas. The best potential targets in terms of decoding accuracy and consistency are located within the depth of and surrounding the lateral fissure bilaterally, such as the (sub)central sulcus, transverse temporal gyrus (Heschls' gyrus), the supramarginal cortex and parts of the insula. These results highlight the potential benefits of extending beyond the motor cortex and reaching the sulcal depth for speech neuroprostheses.