CERs (%) for each method when using RNN-based ASR models with JNAS as the target domain, where [X] indicates an ASR model trained with dataset X and Y + Z indicates a single ASR model trained with datasets Y and Z. [Y]+[Z], on the other hand, represents the integration of the outputs of ASR models trained with datasets Y and Z, respectively. The ROVER method is theoretically applicable to two models, but because majority voting fails and results in a tie, this study reports results only for the case of integration of three models
| ASR model | N | Integration method | ||||
|---|---|---|---|---|---|---|
| w/o LM | SF | DRA | ROVER | Proposed | ||
| [APS] | 1 | 26.3 | 23.9 | 22.4 | – | – |
| [SPS] | 1 | 22.0 | 21.7 | 19.9 | – | – |
| [CV] | 1 | 31.3 | 30.5 | 29.9 | – | – |
| [APS+SPS] | 1 | 18.9 | 16.6 | 15.9 | – | – |
| [APS+SPS+CV] | 1 | 16.0 | 14.5 | 13.7 | – | – |
| [APS]+[SPS] | 2 | 19.4 | 17.9 | – | – | 16.6 |
| [APS]+[SPS]+[CV] | 3 | 19.6 | 18.4 | – | 25.5 | 16.4 |
| Integration method | ||||||
|---|---|---|---|---|---|---|
| w/o | Proposed | |||||
| 1 | 26.3 | 23.9 | 22.4 | – | – | |
| 1 | 22.0 | 21.7 | 19.9 | – | – | |
| 1 | 31.3 | 30.5 | 29.9 | – | – | |
| 1 | 18.9 | 16.6 | 15.9 | – | – | |
| 1 | 16.0 | 14.5 | 13.7 | – | – | |
| 2 | 19.4 | 17.9 | – | – | 16.6 | |
| 3 | 19.6 | 18.4 | – | 25.5 | 16.4 | |
Sharing content requires targeting cookies to be enabled. Please update your cookie preferences to use this feature.