Performance of the recognition model in “cocktail party” situations.
A module is trained on an auditory sentence (“She argues with her sister”) without competing speakers and tested for recognition of this sentence in three conditions: Left column) No competing speaker, Middle column) one competing speaker, and Right column) three competing speakers. Each column shows the second level dynamics, first level dynamics and cochleagram with arbitrary units in neuronal activation. Second level dynamics were successfully reconstructed for the single speaker and also, to an extent, for the speech sample with one competing speaker. In the case of three competing speakers, the module was not able to reconstruct the second level dynamics completely, but showed some signs of recovery at the beginning and at the end of the sentence. Note that the increasing difficulty in reconstruction of the speech message from one to three speakers is not reflected in the prediction errors at the first level (dashed lines), but becomes obvious at the second level.