ISCA Archive SLTU 2018
ISCA Archive SLTU 2018

Low-resource Tibetan Dialect Acoustic Modeling Based on Transfer Learning

Jinghao Yan, Zhiqiang Lv, Shen Huang, Hongzhi Yu

Deep neural network (DNN) based acoustic model has made great breakthroughs in speech recognition. However, lower source Sino-Tibetan languages such as Tibetan still need further studies, especially when dealing with dialects. Based on a TDNN acoustic model trained according to lattice-free MMI criteria, this paper demonstrates baseline systems for two Tibetan dialects: U-Tsang and Amdo. Transfer learning is also employed to improve our systems. Experiment results show that for low resource Tibetan dialect recognition, transfer learning can consistently outperform the baseline.