Children's speech recognition is still a challenging issue. In the case of children's speeches, the accuracy of conventional phrase speech recognition approaches is significantly low. This is mainly owing to the high variability of pronunciation patterns due to children's physical activity. Motivated by this, in this paper, we present a phrase speech recognition system using neural networks. We use a convolutional neural network (CNNs) and its recurrent neural network (RNN) version, say CRNN. Also, both approaches utilize a connectionist temporal classification (CTC) loss function, which allows networks to be trained without any prior alignment. Through experiments using a children's speech database, we show the comparison results of CNN- and CRNN-CTC approaches.