摘要
训练数据和测试数据之间由于信道等差异而引起的不匹配会严重影响语种识别的性能。而在实际应用中,通常只能获得少量的和测试数据匹配的标注数据(目标域数据),以及大量的和测试数据不匹配的标注数据(源域数据)。该文利用迁移学习的方法,通过无监督迁移分量分析(unsupervised transfer component analysis,UTCA),可以合理利用上述两种数据寻找到一个低维子空间,在该空间中,源数据和目标数据之间的分布差异最小,而且数据中有利于分类的属性得以保留,从而提高系统识别性能。实验表明:相对于基线系统,该算法对30s和10s语音的识别性能分别有24.7%和8%的提高。
Distribution mismatches between training and test datasets can greatly reduce the performance of language recognition systems.The mismatch is typically due to variability from changes in the channel and other factors.Real-world applications often have many training samples from other source domains but only a limited number of labeled training samples from the target domain.This study uses transfer learning to find a low-dimensional subspace through unsupervised transfer component analysis（UTCA）.This space minimizes the distribution mismatch between the source and target domain samples while preserving the good data properties.Tests show that the UTCA gives 24.7% and 8% relative improvement at 30 s and 10 s durations over the baseline system.
出处
《清华大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2013年第6期800-803,共4页
Journal of Tsinghua University(Science and Technology)
基金
国家自然科学基金资助项目(61273268
61005019)
关键词
语种识别
迁移学习
迁移分量分析
language recognition
transfer learning
transfer component analysis
