In this paper, we explore how the audio respiration signal can contribute to multimodal analysis of movement qualities. Within this aim, we propose two novel techniques which use the audio respiration signal captured by a standard microphone placed near to mouth and supervised machine learning algorithms. The first approach consists of the classification of a set of acoustic features extracted from exhalations of a person performing fluid or fragmented movements. In the second approach, the intrapersonal synchronization between the respiration and kinetic energy of body movements is used to distinguish the same qualities. First, the value of synchronization between modalities is computed using the Event Synchronization algorithm. Next, a set of features, computed from the value of synchronization, is used as an input to machine learning algorithms. Both approaches were applied to the multimodal corpus composed of short performances by three professionals performing fluid and fragmented movements. The total duration of the corpus is about 17 min. The highest F-score (0.87) for the first approach was obtained for the binary classification task using Support Vector Machines (SVM-LP). The best result for the same task using the second approach was obtained using Naive Bayes algorithm (F-score of 0.72). The results confirm that it is possible to infer information about the movement qualities from respiration audio.

中文翻译：

---

呼吸音频在运动质量多模式分析中的作用

在本文中，我们探讨了音频呼吸信号如何有助于运动质量的多峰分析。在此目标范围内，我们提出了两种新颖的技术，它们使用由靠近嘴巴的标准麦克风捕获的音频呼吸信号和受监督的机器学习算法。第一种方法包括对一组声学特征的分类，这些声学特征是从执行液体运动或肢体运动的人的呼气中提取的。在第二种方法中，呼吸和身体运动的动能之间的人际同步被用来区分相同的素质。首先，使用事件同步算法计算模态之间的同步值。接下来，将从同步值计算出的一组功能用作机器学习算法的输入。两种方法都应用于多模式语料库，该语料库由三名执行流畅和不完整动作的专业人员的短表演组成。语料库的总持续时间约为17分钟。最高的使用支持向量机（SVM-LP）针对二进制分类任务获得了第一种方法的F分数（0.87）。使用朴素贝叶斯算法（F评分为0.72），使用第二种方法获得的相同任务的最佳结果。结果证实，可以从呼吸音频推断出有关运动质量的信息。