Human Perception and Automatic Detection of Speaker Personality and Likability - Influence of Modern Telecommunication Channels

当听到未知的声音时，人类倾向于自发地推断对话者的个性和声音的可爱度。所听到的声音通常是通过通信渠道传输的，例如在基于电话的语音应用中。然而，在以往的人类和自动检测人格特征和讨人喜欢性的研究中，对传递通道效应的研究尚未得到解决。此外，对于这些说话人特征的自动预测，尽管感知评分具有连续性，但主要解决了二值分类任务。该项目将研究不同设置的传输通道，如带宽、编解码器和用户界面，对说话者个性和人类和机器的亲和力检测的影响。将记录拟议分析所需的德语会话语音数据。在人类方面，将采用众包的方式，从大量传输的语音材料中快速可靠地收集听众的评价。在自动方面，回归模型将被考虑用于人格和可爱性预测，采用最先进的技术，如深度神经网络。我们还将研究语音质量指标作为这些说话者特征预测指标的有效性。研究结果将阐明哪些传播渠道可以保留决定感知个性和亲和力的语音属性，以及如何自动预测这些属性。这可用于基于电话语音的应用，其目的是估计可感知的说话者特征并预见随后的用户行为。

When listening to unknown voices, humans tend to make spontaneous inferences about the perceived personality and voice likability of their interlocutors. The voices heard are generally transmitted through communication channels, e.g. in telephone-based speech applications. However, the study of transmission channel effects has not yet been addressed in previous investigations of human and automatic detection of personality traits and likability. Besides, regarding the automatic prediction of these speaker characteristics, the binary classification task has principally been tackled despite the continuous nature of the perceptive ratings. The proposed project will examine the influence of transmission channels of different settings, such as bandwidth, codec and user interface, on speaker personality and likability detection by humans and machines. Conversational speech data in German, needed for the proposed analyses, will be recorded. On the human side, crowdsourcing will be employed to rapidly and reliably gather listeners' assessments from large transmitted speech material. On the automatic side, regression models will be considered for personality and likability prediction, employing state-of-the-art techniques such as deep neural networks. The validity of speech quality measures as predictors of these speaker characteristics will also be studied. The outcomes will elucidate which transmission channels can preserve the voice properties that determine the perceived personality and likability, and how these can be automatically predicted. This can be used in applications based on telephone speech which aim at estimating perceived speaker characteristics and at foreseeing subsequent user behavior.