探索基于CTC的端到端缅甸语语音识别技术
机器学习
2021-05-17 v2 人工智能
摘要
在这项工作中，我们探索了一种基于连接时序分类（Connectionist Temporal Classification, CTC）的端到端自动语音识别（Automatic Speech Recognition, ASR）模型用于缅甸语。我们针对模型拓扑结构进行了一系列实验，其中添加或去除了卷积层，使用了不同深度的双向长短期记忆（bidirectional long short-term memory, BLSTM）层，并研究了不同的标签编码方法。实验在低成本资源场景下使用我们录制的近26小时缅甸语语音语料库进行。最佳模型在测试集上取得了4.72%的字符错误率（character error rate, CER）和12.38%的音节错误率（syllable error rate, SER）。
引用
@article{arxiv.2105.06253,
title = {Exploring CTC Based End-to-End Techniques for Myanmar Speech Recognition},
author = {Khin Me Me Chit and Laet Laet Lin},
journal= {arXiv preprint arXiv:2105.06253},
year = {2021}
}
备注
This is a preprint of the chapter: Chit K.M.M., Lin L.L., Exploring CTC Based End-To-End Techniques for Myanmar Speech Recognition, published in Advances in Intelligent Systems and Computing, vol 1324, edited by Vasant P., Zelinka I., Weber GW., 2021, Springer, Cham reproduced with permission of Springer. The final authenticated version is available at https://doi.org/10.1007/978-3-030-68154-8_87