Humans are remarkably skilled at listening to one speaker inin the ignored speech source, over and above onsets of the mixture and the attended source. This suggest that the auditory cortex initially reconstructs acoustic onsets that could belong to any speech source, even when those onsets are acoustically obscured by another source. The auditory cortex tracked onsets in the acoustic mixture with a lower latency (~70 ms) than obscured onsets (~100 ms), suggesting a neural processing cost to recovery of obscured onsets. Because acoustic onsets precede sustained source-specific information in the acoustic spectrogram, these representations of onsets could provide important cues for a subsequent processing stage, at which features of the attended source are analyzed more selectively.