In this study, the improvement in a new audio effect called sound collage, whereby one sound waveform (target sound) is synthesized using another sound waveform (elemental sound), is investigated. In this study, we propose a new model of convolutional NMF (CNMF) with constraints. And we compared the performance of three methods: the original CNMF, the new CNMF constraint model, and Driedger's NMF (non-negative matrix factorization method). Sound collage sounds are synthesized using a combination of animal calls as the target sound and several instrumental sounds as the elemental sounds. Psychological experiments are conducted to examine the degrees to which the target sound (reproducibility) and instrumental character (instrumentality) are represented. The results confirm that the instrumental nature of the synthesized sounds for both models improve compared with CNMF.