基于主动学习的环境音分类研究

环境音分类是当前语音识别领域的研究热点。主动学习是利用未标记数据,在少量标记数据代价下提高监督学习算法的分类性能的方法。文中提出了熵优先采样（Entropy Priority Sampling,EPS）方法和简单不一致采样（Simple Disagreement Sampling,SDS）方法作为主动学习选择样本的策略。针对环境音数据,提取11维的CELP音频特征,采用单一分类器与EPS,SDS方法对不同标记训练样本比例下的分类实验结果进行了比较分析。结果表明,主动学习方法在标记样本数较少的情况下,能取得较好的分类效果,并且EPS方法的性能优于SDS方法。

Research on Environmental Audio Classification Based on Active Learning

Environmental audio classification has been the focus in the field of speech recognition.Active learning enhances the performance of supervised learning classification under the case of few labeled data.Propose EPS（ Entropy Priority Sampling） and SDS（ Simple Disagreement Sampling） methods as the selecting sampling strategies in active learning.For the given environmental audio data,the CELP features in 11 dimensions are extracted.The experiments with the single classifier,EPS and SDS on the environmental audio are carried out in order to illustrate the results of the proposed methods and compare their performance under different percent training sample.The experimental results show that active learning can effectively improve the performance of environmental audio data classification,even under the fewer number of the training examples.The EPS method outperforms the SDS.