The changes of short-term feature in a musical note can reflect the timbre of musical instrument. In this paper, the short-term features of frames in a musical note instead of in a window of fixed length are integrated into a note feature vector, which captures the dynamic characteristics of musical note, by diverse integration methods such as statistic (ST) model, autoregressive (AR) model and modulation model. For musical instrument classification, it is verified that the pitch-frequency cepstral coefficients (PFCCs) as short-term features are better than the mel-frequency cepstral coefficients (MFCCs). We select the PFCCs and the partials amplitudes as short-term feature to generate note features by integration models respectively. In the experience, the Gaussian mixture model (GMM) classifier and Random Forest (RF) classifier are exploited to testify that these note features are useful for musical instrument classification. Results demonstrate that the combination of ST model, univariate AR model and 2-dimension modulation model is exceptionally superior to the short-term feature whether it is based on partials amplitudes or based on PFCCs.