Since the values of ACD are not consistent with the signal
quality of the converted voice, we originally tune the vales of K and
α by listening to the converted voices. Now, we know that the
measured values of variance ratio (VR) are almost consistent with the
converted-voice quality. Therefore, we experiment to decide the values
of K and α according to the measured VR values. The measured VR
for different K values are listed in Table B whereas the measured VR
for different α values are listed in Table C.
From Table B, it is seen that the average VR is not sensitive to K (beam width for dynamic programming) value within the range, 16 to 32, but reach the maximum when K=24. Therefore, we decide to use K=24.
From Table C, it is seen that the average VR is not sensitive to α (weighting factor)
value within the range, 0.25 to 6. Thus, we cannot but to suspect
whether VR cannot distinguish the converted-voice quality when VR
values are close. Also, we cannot distinguish the converted-voice
qualities when α is varied from 0.5 to 2 by perception.
Therefore, we select to use α=1.5.