Hybrid CTC Language Identification Structure for Mandarin-English Code-Switching ASR
Hengxin Yin, Guangyu Hu, Fei Wang, Pengfei Ren · 2022 13th International Symposium on Chinese Spoken Language Processing (ISCSLP) · 2022
With the advent of globalization, there is a growing demand for code-switching automatic speech recognition (ASR), which can accurately discriminate a speaker who alternates words of two or more languages within a single sentence or across sentences. The common phenomenon of Mandarin-English code-switching ASR is quite common around the world, like China and Singapore. In this paper, we propose a hybrid CTC language identifycation architecture and a complete model training method for Mandarin-English code-switching ASR. We propose an ASR architecture that adds a LID layer after CTC decoder which based on CTC-AED architecture to calculate the CTC LID loss at the frame level. Furthermore, a fusion loss based on CTC loss, Attention loss and CTC LID loss is used to train Mandarin-English code-switching ASR model. In order to enable the ASR model to quickly converge and strengthen the language identification ability, a dynamic CTC LID loss weight is introduced. Results show that the proposed model architecture and model training method can effectively improve performance of the Mandarin-English code-switching ASR model. The final system achieves a Mixed Error Rate (MER) of 20.9% in the ISCSLP2022 Magichub Code-Switching ASR Challenge.