This paper describes a new approach in musical instrument identification, an important task in the field of Music Information Retrieval (MIR). It is based on our previously developed probabilistic model which approximates the input audio spectrogram with a mixture of Gaussians. The EM algorithm is used to estimate the model parameters and calculate our newly proposed Harmonic Temporal Timbre Energy Ratio and Harmonic Temporal Timbre Envelope Similarity features. We then use these features in a novel boosting algorithm to perform the instrument classification. Contrary to traditional boosting methods, like the very popular AdaBoost, our new method uses probabilistic decision-making for hypotheses in each iteration, which results in better noise handing and higher classification accuracy.