TABLE II.

Performance of the $F_0$ estimation algorithms (actual speech signals). The evaluation of the $F_0$ estimation algorithms uses all the 65 actual speech signals, where for each signal we use 90 $F_0$ estimates (thus $N=65\times 90=5850$). The results are in the form mean ± standard deviation. The last four rows are the approaches to combine the outputs of the $F_0$ estimation algorithms using the median from all algorithms, OLS, IRLS, and adaptive KF. The best individual $F_0$ estimation algorithm and the best combination approach are highlighted in bold. The ME in the second column is used to illustrate the bias of each algorithm.

Algorithm	ME (Hz)	MAE (Hz)	MRE (%)	RMSE (Hz)
dypsa	−0.78	14.42 ± 26.32	5.54 ± 8.44	25.86 ± 32.89
praat1	−0.03	29.22 ± 57.23	13.28 ± 24.08	31.67 ± 57.10
praat2	−0.03	29.05 ± 56.86	13.21 ± 24.00	31.47 ± 56.71
rapt	−0.04	28.30 ± 63.47	8.63 ± 17.98	34.21 ± 65.89
shrp	−0.01	18.78 ± 47.77	6.85 ± 16.86	26.91 ± 55.21
swipe	0.10	3.06 ± 7.01	1.18 ± 2.48	6.22 ± 13.46
yin	−0.03	16.36 ± 47.34	6.16 ± 16.32	23.35 ± 51.77
ndf	−0.01	15.12 ± 60.66	4.16 ± 15.24	17.66 ± 60.87
tempo	−0.03	50.67 ± 99.23	17.69 ± 31.08	53.21 ± 100.92
xsx	−0.08	33.43 ± 52.11	16.85 ± 25.90	39.57 ± 56.81
Median	−0.17	18.90 ± 46.27	7.71 ± 18.11	24.71 ± 49.15
OLS	−0.78	4.08 ± 7.76	1.55 ± 2.62	7.58 ± 13.82
IRLS	−0.03	3.17 ± 7.03	1.23 ± 2.49	6.53 ± 13.57
KF	−0.03	2.49 ± 5.04	0.97 ± 1.82	4.95 ± 9.19