Table 1.
Matching (%) of CDP++, RC00 and Sequitur pronunciations of the 915 nonwords from (Mousikou et al., 2017) against human pronunciations from the same work
| Pronunciations | 1st | 2nd | 3rd | 4th | 5th | 6th | 7th | Match | Absent |
|---|---|---|---|---|---|---|---|---|---|
| Strict scoring | |||||||||
| CDP++ | 44 | 16 | 8 | 4 | 2 | 1 | 0 | 77 | 23 |
| RC00 | 55 | 20 | 7 | 4 | 2 | 1 | 1 | 89 | 11 |
| Sequitur | 40 | 15 | 6 | 4 | 2 | 1 | 1 | 70 | 30 |
| Lenient scoring | |||||||||
| CDP++ | 50 | 13 | 7 | 4 | 2 | 1 | 0 | 79 | 21 |
| RC00 | 69 | 14 | 5 | 2 | 1 | 1 | 0 | 92 | 8 |
| Sequitur | 50 | 12 | 5 | 3 | 2 | 1 | 1 | 75 | 25 |
The i-th column reports the percentage of model/algorithm output that matches the i-th most frequent human pronunciation; Match is the percentage of output matching any human pronunciation; Absent is the complement of Match. Strict criteria means that two pronunciations (phoneme strings) have to match exactly, while lenient means that substitution errors between short vowels and schwa (both ways) are forgiven