Automatic Speech Recognition Using Cross-Lingual Pretrained Model and Transfer Learning by Amalia Zahra, Ph.D. Lecturer at Binus University in Jakarta, Indonesia

In speech recognition, it has been a challenge to build a model for under-resourced languages, and Indonesian is one of them. However, with the use of pretrained model and transfer learning, such a challenge can be overcome with much smaller amount of training data. This finding opens further opportunity for our local dialects. Indonesia has approximately 718 local dialects; most of which have not been explored for speech recognition (or other speech processing tasks) due to lack of resources. By using some information contained in a model trained with (richer) datasets of other languages, it is now possible to fill such a gap. It is proven by the fact that our current research achieves 12% of Word Error Rate (WER) using lesser amount of data for Indonesian speech recognition. This approach of using pretrained model and transfer learning is certainly not only applicable to Indonesian or Indonesia’s local dialects, but also other under-resourced languages.