人工智能培训

搜索

人工智能论文:多语言序列到序列语音识别系统的分析(Analysis of Multilingual Sequence-to-Sequence speech re

[复制链接]
jamiezhao 发表于 5 天前 | 显示全部楼层 |阅读模式
jamiezhao 5 天前 84 0 显示全部楼层
人工智能论文:多语言序列到序列语音识别系统的分析(Analysis of Multilingual Sequence-to-Sequence speech recognition systems)本文研究了传统隐马尔可夫模型(HMM)系统中序列到序列(seq2seq)自动语音识别(ASR)中开发的各种多语言方法的应用。在一组由Babel数据组成的数据中,我们首先展示了具有堆叠瓶颈(SBN)特征的多语言训练的有效性。然后,我们探讨了基于CTC注意网络的多语言seq2seq模型的各种体系结构和训练策略,包括输出层,CTC和/ orattention组件重新训练的组合。当目标语言未包含在原始多语言训练数据中时,我们还研究了语言转移学习在非常低资源情况下的有效性。有趣的是,我们发现多语言特征优于多语言模型,这一发现表明我们可以有效地结合通过这些多语言功能技术,使用seq2seq系统实现了HM系统的优势。
This paper investigates the applications of various multilingual approachesdeveloped in conventional hidden Markov model (HMM) systems tosequence-to-sequence (seq2seq) automatic speech recognition (ASR).On a setcomposed of Babel data, we first show the effectiveness of multi-lingualtraining with stacked bottle-neck (SBN) features.Then we explore variousarchitectures and training strategies of multi-lingual seq2seq models based onCTC-attention networks including combinations of output layer, CTC and/orattention component re-training.We also investigate the effectiveness oflanguage-transfer learning in a very low resource scenario when the targetlanguage is not included in the original multi-lingual training data.Interestingly, we found multilingual features superior to multilingual models,and this finding suggests that we can efficiently combinethe benefits of theHMM system with the seq2seq system through these multilingual featuretechniques.人工智能论文:多语言序列到序列语音识别系统的分析(Analysis of Multilingual Sequence-to-Sequence speech recognition systems) x3P3RRvEN1LlevM7.jpg
URL地址:https://arxiv.org/abs/1811.03451     ----pdf下载地址:https://arxiv.org/pdf/1811.03451    ----人工智能论文:多语言序列到序列语音识别系统的分析(Analysis of Multilingual Sequence-to-Sequence speech recognition systems)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

jamiezhao当前离线
新手上路

查看:84 | 回复:0

快速回复 返回顶部 返回列表