ABSTRACT

While voice communication of emotion has been researched for decades, the accuracy of automatic voice emotion recognition (AVER) is yet to improve. In particular, the intergenerational communication has been under-researched, as indicated by the lack of an emotion corpus on child–parent conversations. In this paper, we presented our work of applying Support-Vector Machines (SVMs), established machine learning models, to analyze 20 pairs of child–parent dialogues on everyday life scenarios. Among many issues facing the emerging work of AVER, we explored two critical ones: the methodological issue of optimising its performance against computational costs, and the conceptual issue on the state of emotionally neutral. We used the minimalistic/extended acoustic feature set extracted with OpenSMILE and a small/large set of annotated utterances for building models, and analyzed the prevalence of the class neutral. Results indicated that the bigger the combined sets, the better the training outcomes. Regardless, the classification models yielded modest average recall when applied to the child–parent data, indicating their low generalizability. Implications for improving AVER and its potential uses are drawn.

摘要

虽然情感的语音交流已经研究了几十年，但自动语音情感识别（AVER）的准确性还有待提高。特别是，代际交流的研究不足，儿童与父母对话中缺乏情感语料库就表明了这一点。在本文中，我们介绍了我们应用支持向量机 (SVM)、建立的机器学习模型来分析 20 对日常生活场景中的亲子对话的工作。在 AVER 新兴工作面临的许多问题中，我们探讨了两个关键问题：针对计算成本优化其性能的方法问题，以及情绪中立状态的概念问题。我们使用 OpenSMILE 提取的简约/扩展声学特征集和一组小/大带注释的话语来构建模型，并分析了类中性的普遍性。结果表明，组合集越大，训练效果越好。无论如何，分类模型在应用于亲子数据时产生了适度的平均召回率，表明它们的普遍性较低。绘制了改进 AVER 及其潜在用途的含义。