Abstract
It is a common assumption that the harmonic series determines almost exclusively the timbre of pitched musical instruments. In this paper we test this assumption and then perform automatic musical instrument recognition in polyphonic music using solely the harmonic series for distinguishing among the different sound sources.

We test the significance of the harmonic series by separating all samples in an extensive sound collection into harmonic and non-harmonic components using additive analysis/synthesis and performing instrument recognition separately on the original samples, the harmonic series and the residuals. It is shown that although the harmonic series by itself is indeed sufficient for achieving high instrument recognition rates, the non-harmonic residuals contain considerable timbre information.

Following these results, two polyphonic instrument recognition techniques are introduced which use automatically estimated harmonic series for source differentiating. In Harmonic-Series Resynthesis, instrument recognition is performed on polyphonic music by resynthesizing each note out of its estimated harmonic series. In Source Reduction, the energy of all notes except one is reduced significantly in order to facilitate its instrument recognition.