Table 6 Monolingual ASR model results.
From: Multilingual end-to-end ASR for low-resource Turkic languages with common alphabets
Languages | Details | Total hours | # Uttr-s | Validation set | Test set | ||
|---|---|---|---|---|---|---|---|
CER | WER | CER | WER | ||||
Tatar | s.p.(speed perturbation): 0.9,1.0,1.1 | 29 | train: 20,204, val: 2812 | 4.5 | 17.0 | 7.0 | 22.5 |
Kazakh | s.p.: 0.9,1.0,1.1 | 1 | train: 406, val: 316 | 66.3 | 123.6 | 67.7 | 124.2 |
Sakha (Yakut) | s.p.: 0.9,1.0,1.1 | 6 | train: 1633, val: 1083 | 29.2 | 79.7 | 32.8 | 85.5 |
Bashkir | no s.p | 265 (255) | train: 178,522, val: 14,577 | 1.8 | 6.4 | 1.7 | 6.1 |
Kyrgyz | s.p.: 0.9,1.0,1.1 | 44(6.5 kept) | train: 4010, val: 502 | 17.7 | 54.3 | 17.9 | 55.3 |