Transfer learning for automatic speech recognition systems

摘要

This paper presents the effects of transfer learning on deep neural network based speech recognition systems. The source acoustic model is trained on a large corpus of call centers telephony records, and an acoustically mismatched out-of-domain data that consists of the meeting recordings of the Grand National Assembly of Turkey is selected as the target. Our goal is to adapt the source model to the target data using transfer learning, and we investigate the effects of different target training data sizes, transferred layer counts and feature extractors on transfer learning. Our experiments show that for all target training sizes, the transferred models outperformed the models that are only trained on the target data, and the model that is transferred using 20 hours of target data achieved 7.8% higher recognition accuracy than the source model.