Table 9.
Human segmentation, InaSpeech segmentation, and the proposed improved InaSpeech segmentation are all benchmarked using the E2E − T transformer and the DNN-HMM WER (%) results on two sets: Uzbek Test and Hidden Test.
Uzbek_Test | Hidden_Test | |||
---|---|---|---|---|
E2E − T | DNN-HMM | E2E − T | DNN-HMM | |
HS | 17.6 | 24.4 | 14.4 | 18.1 |
IS | 41.6 | 28.7 | 34.8 | 25.7 |
IMP_IS | 19.7 | 31.1 | 16.5 | 28.3 |