BROWSE
Detail
Conference Paper
A Data-Augmented Transfer Learning Method for the Speech Recognition in Domains with Sparse Speech Data
Cited - time in 
Share 
Share
- Authors
- Byung Ok Kang, Hyeong Bae Jeon
- Issue Date
- 2022-10
- Citation
- International Congress on Acoustics (ICA) 2022, pp.1-5
- Language
- English
- Type
- Conference Paper
- Abstract
- In this paper, we propose a data-augmented transfer learning method for the purpose of improving the performance of speech recognition in the domains with sparse training data where it is difficult to collect large amounts of labeled/unlabeled speech data. As the first step, the proposed method augments speech corpus of acoustic characteristics such as speaker and channel/noise environment similar to that of the target domain using speech corpus of other domains that is relatively easy to collect speech data for training. Next, it performs the proposed transfer learning in the form of combination of self-training and teacher/student learning using source/augmented speech corpus with input. For evaluation, we performed experiments on the AMI corpus task and the call-center speech-to-text (STT) task, and the proposed approach outperformed the existing teacher/student-based transfer learning method.
- KSP Keywords
- AMI corpus, Acoustic characteristics, Call center, Learning methods, Speech corpus, Speech-To-Text(STT), Student Learning, Target domain, Transfer learning, self-training, speech recognition