Efficient Conformer-Based CTC Model for Intelligent Cockpit Speech Recognition
Hanzhi Guo, Yunshu Chen, Xukang Xie, Gaopeng Xu, Wei Guo · 2022 13th International Symposium on Chinese Spoken Language Processing (ISCSLP) · 2022
In this paper, we discuss the rationale of our work for automatic speech recognition (ASR) in the ISCSLP 2022 Intelligent Cockpit Speech Recognition Challenge and provide a detailed description. We proposed a combination of intermediate CTC-based loss regularization(Inter-CTC) with self-conditioned folded encoders, also deployed with TLG attention-rescore-based acoustic model. This combination finds good solutions for Intelligent Cockpit Speech recognition, as it makes accuracy improvements. The Character Error Rate (CER) of the model in Track II (Unlimited model size track) decreased by 38.12% compared to the baseline model, achieving second place in the ranking period with a CER of 9.86%. For Track I (Limited model size track), we apply knowledge distillation to train the Teacher-Student model from the unlimited track, with a combining Track I and Track II Kloss. The CER of the Track II model drops by 40.18% from the baseline model, achieving third place in the ranking period with a CER of 13.39%.