Papers by Gernot Kubin
Conversational Speech Recognition Needs Data? Experiments with Austrian German (2022.lrec-1)
Copied to clipboard
| Challenge: | Recent advances in self-supervision have allowed such scenarios to take advantage of large amounts of otherwise unrelated data. |
| Approach: | They characterise an Austrian German conversational task using a non-pre-trained baseline and a leave-one-conversation out technique. |
| Outcome: | The proposed model fine-tunes well with a non-pre-trained baseline and shows that the advantage of pre-training arises from the larger database rather than the self-supervision. |