This paper studies speech command recognition and proposes a Differential Evolution based search strategy for CNN hyper-parameters. The goal is to reduce manual tuning cost while improving classification quality on the Google Speech Commands dataset.