训练数据有限的英文语音重音标注研究

Stress detection in English sentences with limited training data

大规模语料库的手工韵律标注消耗大量的时间和人力。这篇论文的目的在于研究如何充分利用少量的手工标注数据训练得到尽可能精确的语音重音自动标注器。论文列举并对比了四种训练方法的效果。在训练中结合声学分类器和语言学分类器,同时使用了综合分类器做后期优化。在实验中,使用机器数据训练声学分类器,并将有限的手工数据用于后期综合分类器能得到最佳的标注正确率。最终的正确率达到了94.0%,与手工标注的正确率上限97.2%比较接近。

It is money and labor consuming to label stressed syllables manually,especially when the speech database is very large.An efficient and reliable automatic prosody labeler is always desired.When training data is limited,how to get the best use of it？ This paper proposes the optimization in using training data for automatic stress detection in English speech utterances.The detector consists of a linguistic classifier,an acoustic classifier and an AdaBonst classifier that can improve the accuracy by using more features and manual labels.The best resuh we obtained is 94.0%,which is approaching to the self-agreement ratio （97.2%） of the same annotator,or the upper hound of the performancc.