Neural representation of lexico-semantics in speech processing has been revealed in recent years. However, to date, how the brain makes sense of the higher-level semantic gist (topic keywords) of a continuous speech remains mysterious. Capitalizing on a generative probabilistic topic modelling algorithm on speech materials to which participants listened while their brain activities were recorded by Magnetoencephalography (MEG), here we show spatio-temporal neural representation of topic keywords in a multi-speaker environment where task-relevant (attended) and -irrelevant (unattended) speech co-exits. We report the difference of neural representation between salient and less salient semantic gist of both attended and unattended speech. Moreover, we show that greater sensitivity to semantically salient unattended speech in the left auditory and motor cortices negatively mediates attended speech comprehension.