arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 1901.01239cs.CL
收藏

端到端CTC模型的说话人自适应

Speaker Adaptation for End-to-End CTC Models

Microsoft AI and Research(微软人工智能与研究事业部)

Ke Li, Jinyu Li, Yong Zhao, Kshitiz Kumar, Yifan Gong

摘要:

我们提出了两种方法,在端到端(E2 E)自动语音识别系统的说话人自适应。一个是Kullback-Leibler散度(KLD)正则化,另一个是多任务学习(MTL)。这两种方法都旨在解决E2 E系统中说话人自适应的数据稀疏性,特别是输出目标稀疏性问题。KLD正则化通过迫使自适应模型的输出分布接近未自适应模型来自适应模型。MTL利用联合训练的辅助任务来提高主任务的性能。我们研究了我们的方法E2 E连接主义时间分类(CTC)模型与三种不同类型的输出单元。微软短信听写任务的实验表明,MTL优于KLD正则化。特别是,MTL适应获得了8.8%和4.0%的相对字出错率减少(WERR)的监督和无监督的适应字CTC模型,和9.6%和3.8%的相对WERR的混合单元CTC模型,分别。

英文摘要:

We propose two approaches for speaker adaptation in end-to-end (E2E) automatic speech recognition systems. One is Kullback-Leibler divergence (KLD) regularization and the other is multi-task learning (MTL). Both approaches aim to address the data sparsity especially output target sparsity issue of speaker adaptation in E2E systems. The KLD regularization adapts a model by forcing the output distribution from the adapted model to be close to the unadapted one. The MTL utilizes a jointly trained auxiliary task to improve the performance of the main task. We investigated our approaches on E2E connectionist temporal classification (CTC) models with three different types of output units. Experiments on the Microsoft short message dictation task demonstrated that MTL outperforms KLD regularization. In particular, the MTL adaptation obtained 8.8\% and 4.0\% relative word error rate reductions (WERRs) for supervised and unsupervised adaptations for the word CTC model, and 9.6% and 3.8% relative WERRs for the mix-unit CTC model, respectively.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑