Voice Quality Modelling for Expressive Speech Synthesis

Table 5

The confusion matrix (%) and 𝐹1 measures in the expressive speech styles identification for the reference configurations (“Natural” and “ResHNM”) and HNM transformation configurations (“HNMPro,” “HNMProJiSh,” and “HNMProVoQ”).

(%)	HAP	SEN	AGG	SAD	Others
Natural
HAP	94.1	0.0	5.9	0.0	0.0
SEN	0.0	100.0	0.0	0.0	0.0
AGG	0.0	0.0	100.0	0.0	0.0
SAD	0.0	9.4	0.0	90.6	0.0
(𝐹1)	0.97	0.96	0.97	0.95	—
ResHNM
HAP	87.1	0.0	11.8	0.0	1.2
SEN	1.2	95.3	1.2	2.4	0.0
AGG	1.2	0.0	98.8	0.0	0.0
SAD	0.0	11.8	0.0	88.2	0.0
(𝐹1)	0.92	0.92	0.93	0.93	—
HNMPro
HAP	30.6	3.5	17.6	28.2	20.0
SEN	1.2	31.8	9.4	40.0	17.6
AGG	35.3	1.2	21.2	23.5	18.8
SAD	2.4	18.8	10.6	41.2	27.1
(𝐹1)	0.36	0.41	0.27	0.35	—
HNMProJiSh
HAP	30.6	3.5	16.5	27.1	22.4
SEN	3.5	30.6	4.7	36.5	24.7
AGG	32.9	2.4	18.8	22.4	23.5
SAD	4.7	17.6	1.2	58.8	17.6
(𝐹1)	0.36	0.40	0.27	0.48	—
HNMProVoQ
HAP	34.1	3.5	25.9	14.1	22.4
SEN	0.0	40.0	5.9	34.1	20.0
AGG	23.5	2.4	31.8	16.5	25.9
SAD	3.5	20.0	1.2	64.7	10.6
(𝐹1)	0.42	0.48	0.39	0.56	—