Deep phenotyping of schizophrenia using speech data could significantly further our understanding of the condition, with the potential to revolutionise healthcare for psychosis. A core symptom of schizophrenia is formal thought disorder, manifesting in subtle changes of the patient’s speech which can appear incoherent and disorganized. Graph theoretical tools have been used to measure disorganised syntax in speech transcripts from patients with psychotic disorders. However, these existing graph theoretical tools ignore the semantic content of speech, which has been show to be altered in patients with psychotic disorders. Mapping semantic content of speech as a network and analysing the network using graph theory could be a powerful approach for detailed characterisation of psychosis symptoms. We therefore developed an algorithm that captures the semantic content of transcribed speech as a network. Methods. Our tool uses Natural Language Processing (NLP) to construct speech networks from transcripts of spoken text (e.g. I see a man). Nodes represent entities (e.g. I, man) and edges represent relations between nodes (e.g. see). We have released our tool as a freely available python package, Networks of Transcript Semantics (netts). We used netts to characterise the properties of speech networks from a general public sample and test for group differences in a clinical sample consisting of healthy participants, first episode psychosis patients and subjects at clinical high risk of psychosis. Results. Semantic speech networks from first episode psychosis patients performing a picture description task included more, smaller connected components than those from healthy control subjects, suggesting that the semantic content of speech graphs from patients was more fragmented. The semantic speech networks show robust case-control differences in schizophrenia that are related to symptom severity such that participants with smaller connected components scored higher on the Negative Thought and Language Index scale. A clustering analysis suggested that semantic speech networks captured novel signal not already described by existing NLP measures. Discussion. The semantic speech networks proposed here provide a useful framework for mapping the content of speech in much more detail than previously possible. Because of the richness of information contained in the semantic speech networks, they may also be useful for studying other mental health conditions and use in other fields of research.