Table 6 Word error rates for all four voices for different amounts of data pruning
From: Developing a unit selection voice given audio without corresponding text
Percentage units used | Word error rate (%) | |||
|---|---|---|---|---|
| Â | Test data = Olive | Test data = lecture | ||
| Â | ASR trained on Olive | ASR trained on LibriSpeech | ASR trained on lecture | ASR trained on LibriSpeech |
100 | 14.25 | 17.21 | 22.13 | 28.17 |
| Â | 13.50 | 15.95 | 20.56 | 23.90 |
50 | 8.11 | 9.56 | 16.25 | 17.15 |
30 | 6.26 | 6.28 | 13.25 | 13.87 |