Identification of particular voices in polyphonic and polytimbral music is a task often performed by musicians in their everyday life. However, the automation of this task is very challenging, because of high complexity of audio data. Usually additional information is supplied, and the results are far from satisfactory. In this paper, we focus on classical music recordings, without requiring the user to submit additional information. Our goal is to identify musical instruments playing in short audio frames of polyphonic recordings of classical music. Additionally, we extract pitches (or pitch ranges) which combined with instrument information can be used in score-following and audio alignment, see e.g. [9, 20], or in works towards automatic score extraction, which are a motivation behind this work. Also, since instrument timbre changes with pitch, separate classifiers are trained for various pitch ranges for each instrument. Four instruments are investigated, representing stringed and wind instruments. The influence of adding harmonic (pitch-based) features to the feature set on the results is also investigated. Random forests are applied as a classification tool, and the results are presented and discussed.