Condition Matching / Retrieval
We find that our conditioning input is able to strongly associate generated sound effects to a particular category (Fig. \ref{698992}) allowing good retrieval results for user entered search queries. Judges only marked generated sound effects as not matching at all 24% of the time. We believe these failure cases are due to our model failing to converge on some modalities, and possible bleeding between different modalities due to the gradient penalty present in our training loss\cite{thanh2018catastrophic}. However, we were unable to get our model to converge using the original DCGAN training loss \cite{radford2015unsupervised}. A DCGAN training loss is usually shown to produce better results, in less time, than the WGAN-GP training loss, if training can be stabilized\cite{gulrajani2017improved}. Getting our model to work with a DCGAN training loss may be an interesting avenue for future improvement to the model.