Table 2. Intelligibility evaluation... | Emerald Publishing

Table 2.

Intelligibility evaluation results (CER and WER in %) on speech synthesized from models trained under seven different transcription conditions

Training conditions	Clean		WER-070		WER-162		WER-242		WER-358		WER-437		WER-548
Models / (%) $↓$	CER	WER	CER	WER	CER	WER	CER	WER	CER	WER	CER	WER	CER	WER
Tacotron 2	3.3	7.6	6.1	10.9	7.1	13.1	11.2	19.5	11.5	22.5	17.7	28.7	27.7	45.3
Vits	3.4	6.8	4.1	8.2	5.1	10.3	7.3	14.1	11.9	21.9	17.8	31.6	27.9	46.7
GradTTS	3.4	7.0	3.7	7.6	4.0	8.6	4.3	8.7	5.2	10.6	6.9	14.0	8.9	16.5
Dvt	3.3	6.9	3.6	7.4	4.4	8.9	4.9	9.3	5.4	10.9	6.5	12.8	7.4	13.9
Dvt2	3.2	6.8	3.3	7.4	4.1	8.4	4.2	8.6	4.4	9.6	5.6	10.8	6.9	13.2
Vits-dp	3.1	7.0	3.6	7.6	4.1	8.6	5.9	10.9	8.2	15.2	12.3	22.0	19.3	33.6

Training conditions	Clean		WER-070		WER-162		WER-242		WER-358		WER-437		WER-548
Models / (%) $↓$	CER	WER	CER	WER	CER	WER	CER	WER	CER	WER	CER	WER	CER	WER
Tacotron 2	3.3	7.6	6.1	10.9	7.1	13.1	11.2	19.5	11.5	22.5	17.7	28.7	27.7	45.3
Vits	3.4	6.8	4.1	8.2	5.1	10.3	7.3	14.1	11.9	21.9	17.8	31.6	27.9	46.7
GradTTS	3.4	7.0	3.7	7.6	4.0	8.6	4.3	8.7	5.2	10.6	6.9	14.0	8.9	16.5
Dvt	3.3	6.9	3.6	7.4	4.4	8.9	4.9	9.3	5.4	10.9	6.5	12.8	7.4	13.9
Dvt2	3.2	6.8	3.3	7.4	4.1	8.4	4.2	8.6	4.4	9.6	5.6	10.8	6.9	13.2
Vits-dp	3.1	7.0	3.6	7.6	4.1	8.6	5.9	10.9	8.2	15.2	12.3	22.0	19.3	33.6