Performance degradation caused by conflicting gradient signals when different input types (audio, text) share neural network parameters.