This paper presents a comprehensive description of the Ant multilingual recognition system for the 6th Oriental Language Recognition(OLR 2021) Challenge. Inspired by the transfer learning scheme, the encoder components of language identification(LID) model is initialized from pretrained automatic speech recognition(ASR) networks for integrating the lexical phonetic information into language identification. The ASR model is encoder-decoder networks based on U2++ architecture; then inheriting the shared conformer encoder from pretrained ASR model which is effective at global information capturing and local invariance modeling, the LID model, with an attentive statistical pooling layer and a following linear projection layer added on the encoder, is further finetuned until its optimum. Furthermore, data augmentation, score normalization and model ensemble are good strategies to improve performance indicators, which are investigated and analysed in detail within our paper. In the OLR 2021 Challenge, our submitted systems ranked the top in both tasks 1 and 2 with primary metrics of 0.0025 and 0.0039 respectively, less than 1/3 of the second place, which fully illustrates that our methodologies for multilingual identification are effectual and competitive in real-life scenarios.
Cite as: Lyu, A., Wang, Z., Zhu, H. (2022) Ant Multilingual Recognition System for OLR 2021 Challenge. Proc. Interspeech 2022, 3684-3688, doi: 10.21437/Interspeech.2022-355
@inproceedings{lyu22_interspeech,
author={Anqi Lyu and Zhiming Wang and Huijia Zhu},
title={{Ant Multilingual Recognition System for OLR 2021 Challenge}},
year=2022,
booktitle={Proc. Interspeech 2022},
pages={3684--3688},
doi={10.21437/Interspeech.2022-355}
}