Skip to main content

Advertisement

Figure 2 | EURASIP Journal on Audio, Speech, and Music Processing

Figure 2

From: Percussive/harmonic sound separation by non-negative matrix factorization with smoothness/sparseness constraints

Figure 2

Optimization of β , K SP and K SM parameters using the development database E. (a) β=0, (b) β=0.5, (c) β=1.0, (d) β=1.5 and (e) β=2.0. The number of percussive R p and harmonic R h spectral patterns is equal to 150. It can be observed the effect of the β-divergence cost in the separation process when varying the contribution of the smoothness and sparseness constraints. The goal of this analysis was to find the optimum value of parameters β, KSP and KSM looking for the value of the combination of them that maximizes the SDR measure for the development database E. The maximum SDR value is obtained using β=1.5. This fact may be due to the best trade-off between high and low energy changes in frequency.

Back to article page