Tuning Frequency Estimation Using Circular Statistics

In this document a new approach on tuning frequency estimation based on circular statistics is presented. Two methods are introduced: the calculation of the tuning frequency over an entire audio piece, and the estimation of an adapting reference frequency for a single voice. The results for the tuning frequency estimation look very good for audio pieces where the dominant voices are tuned close to the equal-temperament scale and exhibit only moderate frequency dynamics. For the analysis of popular western music, the method does not achieve very robust results due to the strong frequency dynamics of the human singing voice. Nevertheless, the method could be improved by excluding the singing voice from the calculation taking only the accompaniment into account. The main advantage of the proposed method lies especially in the easy computation of an adaptive reference frequency using an exponential moving average. This adaptive reference can for example be used in the quantization of the singing voice into a note representation.